diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_tensor([5724, 5724])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_tensor([5724, 5724])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..70a231e1fd33478c15c895711bf10797b8171a22 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_tensor([5724, 5724])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera slowly and gracefully advances, capturing a tranquil courtyard steeped in the elegance of traditional Chinese architecture. Ornate crimson lacquer columns, intricately carved with coiling dragons, support the expansive gray-tiled roof with gracefully curved eaves. In the center of the gray slate court, a lone figure performs a classical dance. She is a woman with long, dark hair arranged in an intricate bun, secured with a delicate jade hairpin, her face a serene, focused expression. She wears a flowing, layered silk Hanfu (Chinese Hanfu) in flowing white, its wide sleeves capturing a gentle air in her movements. With poetic precision, she turns slowly and deliberately, one arm gracefully extended, the translucent silk sleeves unfurling like morning mist. Her movements are a study in controlled fluidity, each step and gesture seamlessly blending into the next, embodying a deep sense of inner peace and discipline. The entire scene is bathed in the soft, diffuse light of a tranquil morning, filtering through a thin mist, casting gentle shadows and highlighting the delicate texture of the silk, creating an atmosphere of profound tranquility and contemplation.", + "seed": 12245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_5724_seed_12245_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022228367696516215, + "vlm_logprob": -0.022228367696516215, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.022228367696516215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036684852093458176 + ], + "opposite_logprob_tokens": [ + -6.753668308258057 + ], + "logprob_reward": -0.0036684852093458176, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0036684852093458176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17497283220291138, + 0.053940512239933014, + 0.1046820878982544, + 0.1176915168762207, + 0.08986236900091171, + 0.04042942821979523, + 0.1924004703760147, + 0.046787407249212265, + 0.09761260449886322, + 0.08162074536085129 + ], + "credit_map": [ + 0.022359538823366165, + 0.023053282871842384, + 0.017518743872642517, + 0.023611202836036682, + 0.025228003039956093, + 0.023700932040810585, + 0.01980401575565338, + 0.019697129726409912, + 0.004659032914787531, + 0.010190470144152641, + 0.005010752473026514, + 0.011653170920908451, + 0.005910530686378479, + 0.006702881772071123, + 0.00462610786780715, + 0.005187562201172113, + 0.012262198142707348, + 0.01943179965019226, + 0.010716643184423447, + 0.016764461994171143, + 0.011084700003266335, + 0.0171962883323431, + 0.008212870918214321, + 0.009013128466904163, + 0.00952710211277008, + 0.014337467961013317, + 0.012365644797682762, + 0.020832929760217667, + 0.017440754920244217, + 0.019060621038079262, + 0.01258185226470232, + 0.011545146815478802, + 0.011314626783132553, + 0.010310499928891659, + 0.012304912321269512, + 0.010537814348936081, + 0.015172614715993404, + 0.00991876982152462, + 0.012707426212728024, + 0.007595700211822987, + 0.0053248959593474865, + 0.00558075774461031, + 0.0029906262643635273, + 0.005682244431227446, + 0.004797742702066898, + 0.004865221679210663, + 0.0076508051715791225, + 0.003537133801728487, + 0.022976549342274666, + 0.0185442753136158, + 0.030056491494178772, + 0.0228025633841753, + 0.023418018594384193, + 0.0238812118768692, + 0.02958671934902668, + 0.021134644746780396, + 0.004878699313849211, + 0.009562882594764233, + 0.0035610769409686327, + 0.007466989103704691, + 0.005669798702001572, + 0.006112188566476107, + 0.0068128714337944984, + 0.0027228998951613903, + 0.009484238922595978, + 0.016525214537978172, + 0.00817014928907156, + 0.02120659500360489, + 0.013090063817799091, + 0.007551329210400581, + 0.012328531593084335, + 0.009256487712264061, + 0.007686213590204716, + 0.010688725858926773, + 0.009297824464738369, + 0.011862737126648426, + 0.01132145058363676, + 0.007974447682499886, + 0.009925590828061104, + 0.012863756157457829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0036684852093458176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12063757330179214, + 0.07586464285850525, + 0.09679276496171951, + 0.030496640130877495, + 0.07936511933803558, + 0.03281456232070923, + 0.10162254422903061, + 0.062282562255859375, + 0.06533706188201904, + 0.33478647470474243 + ], + "credit_map": [ + 0.012352355755865574, + 0.012776578776538372, + 0.012378576211631298, + 0.017605720087885857, + 0.017824426293373108, + 0.016116036102175713, + 0.015879744663834572, + 0.015704134479165077, + 0.008720572106540203, + 0.00700397789478302, + 0.008405516855418682, + 0.011305361986160278, + 0.011317619122564793, + 0.009949708357453346, + 0.007762895431369543, + 0.01139899529516697, + 0.010680505074560642, + 0.017015477642416954, + 0.01026975642889738, + 0.019340358674526215, + 0.0103808818385005, + 0.013148064725100994, + 0.00800003856420517, + 0.007957684807479382, + 0.0027042552828788757, + 0.0043989247642457485, + 0.0036631536204367876, + 0.0057206894271075726, + 0.0032498843502253294, + 0.00621489342302084, + 0.0020205022301524878, + 0.00252433936111629, + 0.011116343550384045, + 0.013811584562063217, + 0.007322599180042744, + 0.010954936034977436, + 0.009934690780937672, + 0.008277409709990025, + 0.008954325690865517, + 0.008993232622742653, + 0.003104966366663575, + 0.0059216441586613655, + 0.0028738908004015684, + 0.0047023906372487545, + 0.004183347336947918, + 0.0033240453340113163, + 0.004645569249987602, + 0.004058710299432278, + 0.013701056130230427, + 0.011345349252223969, + 0.010817506350576878, + 0.01013275794684887, + 0.017759213224053383, + 0.012539454735815525, + 0.015385511331260204, + 0.009941707365214825, + 0.012975465506315231, + 0.00555469049140811, + 0.011819359846413136, + 0.005676641594618559, + 0.006489063613116741, + 0.008286954835057259, + 0.004746729042381048, + 0.006733664311468601, + 0.0075708976946771145, + 0.005790283437818289, + 0.0077919261530041695, + 0.007115252315998077, + 0.009370331652462482, + 0.0070389267057180405, + 0.006557499524205923, + 0.014101939275860786, + 0.04787861555814743, + 0.02594611421227455, + 0.03950604796409607, + 0.03658651188015938, + 0.03146607428789139, + 0.0660230964422226, + 0.04182511568069458, + 0.045554935932159424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.01818566583096981 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14563415944576263, + 0.048357900232076645, + 0.08166515827178955, + 0.13885626196861267, + 0.09484744817018509, + 0.07852967083454132, + 0.12672463059425354, + 0.1430574357509613, + 0.07022040337324142, + 0.07210691273212433 + ], + "credit_map": [ + 0.013887220062315464, + 0.03220568969845772, + 0.009851615875959396, + 0.021132899448275566, + 0.018089119344949722, + 0.016669413074851036, + 0.013875656761229038, + 0.019922543317079544, + 0.005207631271332502, + 0.006417654920369387, + 0.004577476531267166, + 0.008830969221889973, + 0.006244393531233072, + 0.006079003214836121, + 0.004424580838531256, + 0.006576187442988157, + 0.009244382381439209, + 0.016379984095692635, + 0.008835426531732082, + 0.01300009898841381, + 0.009048204869031906, + 0.0103310476988554, + 0.007230123970657587, + 0.007595888804644346, + 0.013655525632202625, + 0.017376860603690147, + 0.013340525329113007, + 0.02593165449798107, + 0.020878948271274567, + 0.022575179114937782, + 0.013352255336940289, + 0.011745310388505459, + 0.008823961950838566, + 0.01545739732682705, + 0.012231241911649704, + 0.013379446230828762, + 0.00998405646532774, + 0.010445864871144295, + 0.017678312957286835, + 0.00684716273099184, + 0.007282899226993322, + 0.01435753796249628, + 0.0058843474835157394, + 0.014904100447893143, + 0.009005214087665081, + 0.009911899454891682, + 0.011613966897130013, + 0.0055697085335850716, + 0.02337799407541752, + 0.010716445744037628, + 0.018369285389780998, + 0.01173392217606306, + 0.019617533311247826, + 0.017189158126711845, + 0.009655949659645557, + 0.016064351424574852, + 0.021171584725379944, + 0.02297758311033249, + 0.019236821681261063, + 0.016921553760766983, + 0.011268636211752892, + 0.01575041562318802, + 0.023663438856601715, + 0.012067408300936222, + 0.006806225515902042, + 0.010977668687701225, + 0.006325602065771818, + 0.009475287981331348, + 0.012956976890563965, + 0.006442943587899208, + 0.009064475074410439, + 0.008171225897967815, + 0.008968339301645756, + 0.007967722602188587, + 0.007691647857427597, + 0.00919258687645197, + 0.010408219881355762, + 0.010209846310317516, + 0.009179061278700829, + 0.008489489555358887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0009190387791022658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204335391521454 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204335391521454, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204335391521454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08924917131662369, + 0.04491964727640152, + 0.06397891789674759, + 0.0846899226307869, + 0.07080236077308655, + 0.06680530309677124, + 0.336329847574234, + 0.040368445217609406, + 0.14980672299861908, + 0.05304963141679764 + ], + "credit_map": [ + 0.010402324609458447, + 0.017797723412513733, + 0.007523158099502325, + 0.010560314171016216, + 0.011060807853937149, + 0.010113388299942017, + 0.00901851523667574, + 0.012772941030561924, + 0.005794318858534098, + 0.00544900493696332, + 0.0030002135317772627, + 0.006634197197854519, + 0.00664416654035449, + 0.005805734544992447, + 0.004353580996394157, + 0.007238435093313456, + 0.009350067004561424, + 0.007952640764415264, + 0.006689892616122961, + 0.008955744095146656, + 0.007945975288748741, + 0.009378867223858833, + 0.006063259672373533, + 0.007642475888133049, + 0.007344003766775131, + 0.009452614933252335, + 0.00950287003070116, + 0.017675276845693588, + 0.007240407168865204, + 0.019149204716086388, + 0.008370260708034039, + 0.005955291911959648, + 0.007282032631337643, + 0.00914943590760231, + 0.007009977474808693, + 0.012696253135800362, + 0.011580695398151875, + 0.007642866112291813, + 0.010255738161504269, + 0.005185368005186319, + 0.005064184311777353, + 0.010974554345011711, + 0.005369811784476042, + 0.008822362869977951, + 0.00751135777682066, + 0.007293809670954943, + 0.014782487414777279, + 0.006986740045249462, + 0.04436257481575012, + 0.026556458324193954, + 0.04120470955967903, + 0.03528202697634697, + 0.07231906056404114, + 0.042114321142435074, + 0.03953583166003227, + 0.03495489060878754, + 0.003776273224502802, + 0.009783370420336723, + 0.003183478256687522, + 0.005550772417336702, + 0.003961870912462473, + 0.00473407469689846, + 0.00529119186103344, + 0.004087412264198065, + 0.011185898445546627, + 0.03383100777864456, + 0.00851274374872446, + 0.03456295654177666, + 0.01141511183232069, + 0.009927471168339252, + 0.022446921095252037, + 0.017924638465046883, + 0.004456106573343277, + 0.00923070963472128, + 0.0058555384166538715, + 0.006909890100359917, + 0.0092721376568079, + 0.004565141163766384, + 0.006840532645583153, + 0.0059195831418037415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.06204335391521454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026325294747948647 + ], + "opposite_logprob_tokens": [ + -3.651325225830078 + ], + "logprob_reward": -0.026325294747948647, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026325294747948647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10331990569829941, + 0.032496169209480286, + 0.08260203152894974, + 0.08476123958826065, + 0.3178803622722626, + 0.029737571254372597, + 0.16377799212932587, + 0.05009086802601814, + 0.07554405182600021, + 0.059789836406707764 + ], + "credit_map": [ + 0.01161180343478918, + 0.0194032434374094, + 0.006703884340822697, + 0.014406990259885788, + 0.012444345280528069, + 0.012224367819726467, + 0.014632489532232285, + 0.01189278345555067, + 0.004377563949674368, + 0.004693516064435244, + 0.0023210688959807158, + 0.005337228532880545, + 0.004788060672581196, + 0.004098596982657909, + 0.0031390837393701077, + 0.0037410478107631207, + 0.011008694767951965, + 0.009327352046966553, + 0.008039550855755806, + 0.009705895557999611, + 0.015707876533269882, + 0.011719669215381145, + 0.008557028137147427, + 0.008535969071090221, + 0.008860594592988491, + 0.006617473904043436, + 0.004856178071349859, + 0.019755693152546883, + 0.00526681961491704, + 0.02510223723948002, + 0.007523251231759787, + 0.006778990384191275, + 0.054681528359651566, + 0.03196721896529198, + 0.025300605222582817, + 0.035606954246759415, + 0.05409694090485573, + 0.04071947559714317, + 0.03046564757823944, + 0.04504198953509331, + 0.0034381747245788574, + 0.004799945279955864, + 0.0025293887592852116, + 0.005072970874607563, + 0.0029569941107183695, + 0.0037032298278063536, + 0.004690857604146004, + 0.0025460103061050177, + 0.022998489439487457, + 0.01387651264667511, + 0.027513787150382996, + 0.01836315728724003, + 0.022297216579318047, + 0.01979965530335903, + 0.02036331407725811, + 0.018565857782959938, + 0.004232719540596008, + 0.00770393293350935, + 0.004366009030491114, + 0.006745779421180487, + 0.007024185732007027, + 0.005563159007579088, + 0.010009962134063244, + 0.004445117432624102, + 0.007204823195934296, + 0.006404528394341469, + 0.0072601670399308205, + 0.012966186739504337, + 0.00648100208491087, + 0.009554075077176094, + 0.008691557683050632, + 0.016981707885861397, + 0.005389023572206497, + 0.00793424341827631, + 0.006332747638225555, + 0.010703722946345806, + 0.008266735821962357, + 0.007062492426484823, + 0.007687993347644806, + 0.006412871647626162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.026325294747948647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and gracefully advances, capturing a tranquil courtyard steeped in the elegance of traditional Chinese architecture. Ornate crimson lacquer columns, intricately carved with coiling dragons, support the expansive gray-tiled roof with gracefully curved eaves. In the center of the gray slate court, a lone figure performs a classical dance. She is a woman with long, dark hair arranged in an intricate bun, secured with a delicate jade hairpin, her face a serene, focused expression. She wears a flowing, layered silk Hanfu (Chinese Hanfu) in flowing white, its wide sleeves capturing a gentle air in her movements. With poetic precision, she turns slowly and deliberately, one arm gracefully extended, the translucent silk sleeves unfurling like morning mist. Her movements are a study in controlled fluidity, each step and gesture seamlessly blending into the next, embodying a deep sense of inner peace and discipline. The entire scene is bathed in the soft, diffuse light of a tranquil morning, filtering through a thin mist, casting gentle shadows and highlighting the delicate texture of the silk, creating an atmosphere of profound tranquility and contemplation.", + "seed": 12246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_5724_seed_12246_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04896850306540727, + "vlm_logprob": -0.04896850306540727, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04896850306540727, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01137533225119114 + ], + "opposite_logprob_tokens": [ + -5.386375427246094 + ], + "logprob_reward": -0.01137533225119114, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01137533225119114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09277588874101639, + 0.057338692247867584, + 0.021189535036683083, + 0.01637890376150608, + 0.07894296199083328, + 0.12185150384902954, + 0.04375999793410301, + 0.4594898819923401, + 0.08266148716211319, + 0.025611229240894318 + ], + "credit_map": [ + 0.009461252018809319, + 0.010101448744535446, + 0.017343562096357346, + 0.019619371742010117, + 0.008409072645008564, + 0.006506734527647495, + 0.010816249996423721, + 0.010518190450966358, + 0.003387692617252469, + 0.003250611713156104, + 0.007076164707541466, + 0.00996569637209177, + 0.004203994758427143, + 0.0077690682373940945, + 0.01179021131247282, + 0.009895249269902706, + 0.0018299153307452798, + 0.0017698676092550159, + 0.0024773594923317432, + 0.0049505531787872314, + 0.0014459531521424651, + 0.0016567781567573547, + 0.00462364824488759, + 0.002435458591207862, + 0.0013429077807813883, + 0.0012986671645194292, + 0.002105705440044403, + 0.0032797043677419424, + 0.0015572819393128157, + 0.0013725425815209746, + 0.003171777119860053, + 0.0022503172513097525, + 0.005811229348182678, + 0.004939176142215729, + 0.018139036372303963, + 0.010574114508926868, + 0.007383811753243208, + 0.00889346282929182, + 0.010182281956076622, + 0.013019852340221405, + 0.007784958463162184, + 0.012707527726888657, + 0.014359631575644016, + 0.022555088624358177, + 0.016416402533650398, + 0.014894451946020126, + 0.012819256633520126, + 0.020314183086156845, + 0.004306351765990257, + 0.004192835185676813, + 0.0058587826788425446, + 0.01160674262791872, + 0.0038793315179646015, + 0.0034842807799577713, + 0.003705086652189493, + 0.006726584397256374, + 0.05249829962849617, + 0.018278243020176888, + 0.038982849568128586, + 0.0266854390501976, + 0.027841337025165558, + 0.018496952950954437, + 0.15319480001926422, + 0.12351193279027939, + 0.008874386548995972, + 0.00713999243453145, + 0.008863095194101334, + 0.007208692375570536, + 0.007597058545798063, + 0.006005055736750364, + 0.01708136685192585, + 0.01989184133708477, + 0.0017343407962471247, + 0.0027895220555365086, + 0.00403007585555315, + 0.00566511694341898, + 0.0016473758732900023, + 0.001820857753045857, + 0.0035091829486191273, + 0.004414758179336786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.01137533225119114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307299107313156 + ], + "opposite_logprob_tokens": [ + -3.1680729389190674 + ], + "logprob_reward": -0.04307299107313156, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307299107313156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1096695214509964, + 0.0456036701798439, + 0.017685430124402046, + 0.027260474860668182, + 0.0455048605799675, + 0.08095606416463852, + 0.04995756968855858, + 0.3671613335609436, + 0.23670311272144318, + 0.019497985020279884 + ], + "credit_map": [ + 0.019282400608062744, + 0.014562497846782207, + 0.017178410664200783, + 0.018070010468363762, + 0.011059015989303589, + 0.007826180197298527, + 0.011385990306735039, + 0.010304993949830532, + 0.004252091981470585, + 0.004020362161099911, + 0.005994250997900963, + 0.00865753274410963, + 0.0033223044592887163, + 0.003935348708182573, + 0.008942211046814919, + 0.006479565519839525, + 0.0031509380787611008, + 0.0017414550529792905, + 0.001830312074162066, + 0.0027461510617285967, + 0.0015462636947631836, + 0.0019793901592493057, + 0.0023674804251641035, + 0.002323439111933112, + 0.0036240359768271446, + 0.003360587637871504, + 0.0035427447874099016, + 0.0037647755816578865, + 0.003630648832768202, + 0.002999169286340475, + 0.0031326087191700935, + 0.003205901011824608, + 0.005697288550436497, + 0.004216933622956276, + 0.005991769023239613, + 0.0056055583991110325, + 0.006166810169816017, + 0.004155315924435854, + 0.00852447934448719, + 0.005146698094904423, + 0.007580678444355726, + 0.008384485729038715, + 0.013775376603007317, + 0.01137013640254736, + 0.011637065559625626, + 0.007127817720174789, + 0.0074546802788972855, + 0.013625815510749817, + 0.004381256643682718, + 0.004069020505994558, + 0.006539372261613607, + 0.01286847423762083, + 0.004442387726157904, + 0.005001844372600317, + 0.004892702214419842, + 0.007762507535517216, + 0.041384171694517136, + 0.016035009175539017, + 0.035683561116456985, + 0.024149136617779732, + 0.02173645794391632, + 0.014771618880331516, + 0.11662326753139496, + 0.09677808731794357, + 0.021140852943062782, + 0.01758739911019802, + 0.020293526351451874, + 0.016986748203635216, + 0.017440922558307648, + 0.01429957989603281, + 0.060839831829071045, + 0.06811422854661942, + 0.0014232387766242027, + 0.002397553762421012, + 0.0026611643843352795, + 0.003778396639972925, + 0.0014072861522436142, + 0.0016627570148557425, + 0.002862641355022788, + 0.003304945072159171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04307299107313156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107191201299429 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.01107191201299429, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107191201299429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08948666602373123, + 0.12292752414941788, + 0.05880707502365112, + 0.1383419930934906, + 0.05948905274271965, + 0.22904683649539948, + 0.039436716586351395, + 0.13368293642997742, + 0.0926678329706192, + 0.03611338138580322 + ], + "credit_map": [ + 0.008661412633955479, + 0.008681736886501312, + 0.01763753965497017, + 0.012363085523247719, + 0.007611909881234169, + 0.007598410826176405, + 0.011766472831368446, + 0.015166100114583969, + 0.010506689548492432, + 0.007707152049988508, + 0.01729665882885456, + 0.02320684865117073, + 0.007552274968475103, + 0.009318867698311806, + 0.030524373054504395, + 0.016814671456813812, + 0.003860331838950515, + 0.0037169773131608963, + 0.007463108282536268, + 0.01207861490547657, + 0.0033891291823238134, + 0.0031806856859475374, + 0.017895830795168877, + 0.007222396321594715, + 0.009416905231773853, + 0.013296295888721943, + 0.028055459260940552, + 0.03227763995528221, + 0.007089887745678425, + 0.007476308383047581, + 0.024278821423649788, + 0.016450688242912292, + 0.0065688686445355415, + 0.007346420083194971, + 0.010736013762652874, + 0.007821463979780674, + 0.005400380585342646, + 0.005645242985337973, + 0.00644621392711997, + 0.009524451568722725, + 0.013174853287637234, + 0.026877032592892647, + 0.02382548712193966, + 0.03469477593898773, + 0.04356282204389572, + 0.02075282111763954, + 0.02476886846125126, + 0.04139018803834915, + 0.004129732493311167, + 0.004398344084620476, + 0.004742083605378866, + 0.008283747360110283, + 0.0034656787756830454, + 0.00319122476503253, + 0.003967423923313618, + 0.0072584813460707664, + 0.01504903007298708, + 0.006030035670846701, + 0.013312906958162785, + 0.00862381886690855, + 0.008432733826339245, + 0.005066161043941975, + 0.043136656284332275, + 0.034031592309474945, + 0.005907525308430195, + 0.0026641148142516613, + 0.005875065922737122, + 0.007086485158652067, + 0.005033422727137804, + 0.004574701655656099, + 0.028489500284194946, + 0.03303702175617218, + 0.0023379242047667503, + 0.0038069994188845158, + 0.0037808886263519526, + 0.008590797893702984, + 0.0022532460279762745, + 0.0022881345357745886, + 0.005554135423153639, + 0.007501259446144104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.01107191201299429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036434233188629 + ], + "opposite_logprob_tokens": [ + -2.3503644466400146 + ], + "logprob_reward": -0.10036434233188629, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036434233188629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1176481544971466, + 0.07087220251560211, + 0.05381237342953682, + 0.04388078674674034, + 0.1669527143239975, + 0.14235439896583557, + 0.035171445459127426, + 0.22448213398456573, + 0.10181199014186859, + 0.04301375895738602 + ], + "credit_map": [ + 0.013776302337646484, + 0.013608329929411411, + 0.018174991011619568, + 0.022417375817894936, + 0.011575241573154926, + 0.012184360064566135, + 0.01215868629515171, + 0.013752874918282032, + 0.00460651982575655, + 0.0038687540218234062, + 0.012653824873268604, + 0.012187061831355095, + 0.004156503826379776, + 0.008245492354035378, + 0.015268263407051563, + 0.009885789826512337, + 0.0033924849703907967, + 0.004934537690132856, + 0.006601512897759676, + 0.012209407985210419, + 0.003460000269114971, + 0.0031817834824323654, + 0.01242707297205925, + 0.007605576887726784, + 0.003921369090676308, + 0.005402314476668835, + 0.006683075334876776, + 0.008150716312229633, + 0.0028243092820048332, + 0.003065108321607113, + 0.005804605782032013, + 0.008029296062886715, + 0.01750289648771286, + 0.020507309585809708, + 0.03122641146183014, + 0.02128129079937935, + 0.024760620668530464, + 0.008294775150716305, + 0.020775943994522095, + 0.022603483870625496, + 0.008279703557491302, + 0.015515320003032684, + 0.012764756567776203, + 0.03258940577507019, + 0.025617530569434166, + 0.009894676506519318, + 0.014641592279076576, + 0.023051436990499496, + 0.0027502693701535463, + 0.0034635700285434723, + 0.0038358482997864485, + 0.005880195181816816, + 0.00406981585547328, + 0.00480989133939147, + 0.003576358314603567, + 0.0067855012603104115, + 0.024591322988271713, + 0.010472721420228481, + 0.020008277148008347, + 0.01934206485748291, + 0.013811320997774601, + 0.008520952425897121, + 0.06945069879293442, + 0.05828479304909706, + 0.007225664332509041, + 0.0039627887308597565, + 0.004299933556467295, + 0.006101895123720169, + 0.005605457816272974, + 0.005290177650749683, + 0.03262462466955185, + 0.03670146316289902, + 0.0037575657479465008, + 0.004445453640073538, + 0.005695244763046503, + 0.008216030895709991, + 0.0026871785521507263, + 0.0033351953607052565, + 0.006090008188039064, + 0.00878708716481924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.10036434233188629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789579376578331 + ], + "opposite_logprob_tokens": [ + -2.578958034515381 + ], + "logprob_reward": -0.0789579376578331, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789579376578331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12920749187469482, + 0.07407205551862717, + 0.028094058856368065, + 0.024825001135468483, + 0.048257436603307724, + 0.1574304699897766, + 0.0675329640507698, + 0.2916041910648346, + 0.16245022416114807, + 0.016526062041521072 + ], + "credit_map": [ + 0.013723180629312992, + 0.020234644412994385, + 0.016820935532450676, + 0.026255881413817406, + 0.010949840769171715, + 0.011886557564139366, + 0.016734734177589417, + 0.012601735070347786, + 0.005013674031943083, + 0.004479549825191498, + 0.013075001537799835, + 0.012972992844879627, + 0.0035631090868264437, + 0.007323236670345068, + 0.020389100536704063, + 0.007255402859300375, + 0.003256273688748479, + 0.002288519637659192, + 0.002962200902402401, + 0.0063756126910448074, + 0.0018822604324668646, + 0.001649573096074164, + 0.0055160862393677235, + 0.004163532517850399, + 0.002285224152728915, + 0.0019101478392258286, + 0.0032251127995550632, + 0.00569957634434104, + 0.0014632466481998563, + 0.002486592158675194, + 0.0040804422460496426, + 0.0036746610421687365, + 0.004050116986036301, + 0.003301572985947132, + 0.00809232983738184, + 0.005935191176831722, + 0.007508907932788134, + 0.00679260166361928, + 0.0054756877943873405, + 0.007101031020283699, + 0.018544571474194527, + 0.017173992469906807, + 0.016861045733094215, + 0.028956115245819092, + 0.022718729451298714, + 0.01474762987345457, + 0.014369968324899673, + 0.024058440700173378, + 0.004952230956405401, + 0.005081051029264927, + 0.009834579192101955, + 0.015229419805109501, + 0.005906675010919571, + 0.007621570955961943, + 0.006649147253483534, + 0.01225829403847456, + 0.033921170979738235, + 0.011184156872332096, + 0.02666974999010563, + 0.020426124334335327, + 0.0177745521068573, + 0.011922750622034073, + 0.09476882219314575, + 0.0749368891119957, + 0.012076536193490028, + 0.006960524711757898, + 0.008779508993029594, + 0.00907774269580841, + 0.009300651028752327, + 0.008681348524987698, + 0.051034845411777496, + 0.05653909593820572, + 0.00137407670263201, + 0.0018649890553206205, + 0.001899621682241559, + 0.0031986890826374292, + 0.0014477437362074852, + 0.00138382229488343, + 0.002707088366150856, + 0.002650031354278326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0789579376578331 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and gracefully advances, capturing a tranquil courtyard steeped in the elegance of traditional Chinese architecture. Ornate crimson lacquer columns, intricately carved with coiling dragons, support the expansive gray-tiled roof with gracefully curved eaves. In the center of the gray slate court, a lone figure performs a classical dance. She is a woman with long, dark hair arranged in an intricate bun, secured with a delicate jade hairpin, her face a serene, focused expression. She wears a flowing, layered silk Hanfu (Chinese Hanfu) in flowing white, its wide sleeves capturing a gentle air in her movements. With poetic precision, she turns slowly and deliberately, one arm gracefully extended, the translucent silk sleeves unfurling like morning mist. Her movements are a study in controlled fluidity, each step and gesture seamlessly blending into the next, embodying a deep sense of inner peace and discipline. The entire scene is bathed in the soft, diffuse light of a tranquil morning, filtering through a thin mist, casting gentle shadows and highlighting the delicate texture of the silk, creating an atmosphere of profound tranquility and contemplation.", + "seed": 12247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_5724_seed_12247_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021412408724427224, + "vlm_logprob": -0.021412408724427224, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.021412408724427224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003049131017178297 + ], + "opposite_logprob_tokens": [ + -6.378048896789551 + ], + "logprob_reward": -0.003049131017178297, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.003049131017178297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26516684889793396, + 0.08859439194202423, + 0.070240318775177, + 0.11137951910495758, + 0.2138867825269699, + 0.04331335052847862, + 0.04081769287586212, + 0.06418858468532562, + 0.04109504818916321, + 0.06131744012236595 + ], + "credit_map": [ + 0.03525150567293167, + 0.011981247924268246, + 0.027185659855604172, + 0.04135061055421829, + 0.02067582495510578, + 0.020758258178830147, + 0.057187508791685104, + 0.05077623575925827, + 0.01007242314517498, + 0.011389653198421001, + 0.010492304340004921, + 0.01312995981425047, + 0.009629553183913231, + 0.011675579473376274, + 0.013391930609941483, + 0.008812990970909595, + 0.0056406185030937195, + 0.0068492889404296875, + 0.01752524822950363, + 0.009620113298296928, + 0.007241299841552973, + 0.004706086590886116, + 0.012208993546664715, + 0.006448680069297552, + 0.010267522186040878, + 0.0057224975898861885, + 0.03786676749587059, + 0.01112429704517126, + 0.009312848560512066, + 0.004919387865811586, + 0.02501608058810234, + 0.007150128949433565, + 0.02219395712018013, + 0.007227852009236813, + 0.07285065948963165, + 0.021411949768662453, + 0.011807404458522797, + 0.007169452495872974, + 0.05595814064145088, + 0.015267380513250828, + 0.003309499239549041, + 0.006884077563881874, + 0.006425012368708849, + 0.006435859482735395, + 0.005969033110886812, + 0.003925641532987356, + 0.005650045350193977, + 0.004714184440672398, + 0.0030801957473158836, + 0.0029000509530305862, + 0.005217288620769978, + 0.006874946877360344, + 0.002999278251081705, + 0.005897482857108116, + 0.00843899417668581, + 0.005409458186477423, + 0.004800429567694664, + 0.008941933512687683, + 0.007286087144166231, + 0.010228019207715988, + 0.007520404644310474, + 0.007684167008846998, + 0.009626599960029125, + 0.008100943639874458, + 0.004894979763776064, + 0.0034757861867547035, + 0.007691275794059038, + 0.007764734793454409, + 0.003315106499940157, + 0.003935739863663912, + 0.00583219388499856, + 0.0041852304711937904, + 0.007340097799897194, + 0.005771360360085964, + 0.014044425450265408, + 0.008890648372471333, + 0.00461358530446887, + 0.005004995968192816, + 0.010442459024488926, + 0.0052098766900599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.003049131017178297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070010140538216 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016070010140538216, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070010140538216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39558619260787964, + 0.107048898935318, + 0.0787227675318718, + 0.07912509143352509, + 0.15188021957874298, + 0.0515710823237896, + 0.04734678566455841, + 0.03714773803949356, + 0.029085181653499603, + 0.02248597890138626 + ], + "credit_map": [ + 0.037552908062934875, + 0.019666137173771858, + 0.04858175292611122, + 0.10385233908891678, + 0.03238357976078987, + 0.02194536104798317, + 0.09467185288667679, + 0.03693227842450142, + 0.010016325861215591, + 0.008651903830468655, + 0.014402336440980434, + 0.0246581993997097, + 0.00907176360487938, + 0.011144690215587616, + 0.016802433878183365, + 0.012301255948841572, + 0.005957667250186205, + 0.006543731316924095, + 0.02459586411714554, + 0.010857404209673405, + 0.005109919235110283, + 0.0035440886858850718, + 0.017242910340428352, + 0.004871190059930086, + 0.00993057619780302, + 0.00789915956556797, + 0.012025021016597748, + 0.009277724660933018, + 0.006113936193287373, + 0.007262136321514845, + 0.010632814839482307, + 0.01598373055458069, + 0.014376929961144924, + 0.008434426039457321, + 0.05023231357336044, + 0.016006313264369965, + 0.01058888342231512, + 0.007243710570037365, + 0.03322746977210045, + 0.011770178563892841, + 0.004578093066811562, + 0.005587897263467312, + 0.006747403647750616, + 0.009166412986814976, + 0.005709932651370764, + 0.005821574479341507, + 0.006746799219399691, + 0.007212976925075054, + 0.0032636653631925583, + 0.004123222082853317, + 0.0056760720908641815, + 0.007906840182840824, + 0.003209470072761178, + 0.003622803371399641, + 0.013399139046669006, + 0.006145579274743795, + 0.0035398525651544333, + 0.0030657874885946512, + 0.0037902588956058025, + 0.005411230027675629, + 0.0038600668776780367, + 0.005423424765467644, + 0.0063959164544939995, + 0.005661205388605595, + 0.002973766066133976, + 0.002764520002529025, + 0.005050872918218374, + 0.004679508041590452, + 0.003548880573362112, + 0.0023369721602648497, + 0.0042733727023005486, + 0.0034572938457131386, + 0.0024303216487169266, + 0.002377426018938422, + 0.003936790861189365, + 0.0033685846719890833, + 0.002243814989924431, + 0.0024176170118153095, + 0.003403928829357028, + 0.0023074971977621317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.016070010140538216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26666486263275146, + 0.03970973938703537, + 0.09121820330619812, + 0.03649032115936279, + 0.40920037031173706, + 0.012575525790452957, + 0.03594709932804108, + 0.027573945000767708, + 0.038528863340616226, + 0.04209104925394058 + ], + "credit_map": [ + 0.026882974430918694, + 0.013496781699359417, + 0.030422817915678024, + 0.06701633334159851, + 0.022166717797517776, + 0.013795639388263226, + 0.07113432884216309, + 0.021749278530478477, + 0.0029021285008639097, + 0.004185718949884176, + 0.004101894795894623, + 0.010005190968513489, + 0.003924237098544836, + 0.004284621682018042, + 0.006322633475065231, + 0.003983315546065569, + 0.00700518861413002, + 0.0029973387718200684, + 0.03464231267571449, + 0.013712218962609768, + 0.003954761661589146, + 0.002064085565507412, + 0.022518914192914963, + 0.0043233903124928474, + 0.0033817640505731106, + 0.0023601995781064034, + 0.01261543482542038, + 0.0037979918997734785, + 0.0029336404986679554, + 0.0015427757753059268, + 0.007778494618833065, + 0.0020800214260816574, + 0.04050205647945404, + 0.017874233424663544, + 0.1397561877965927, + 0.04296319559216499, + 0.024300474673509598, + 0.016264449805021286, + 0.09987446665763855, + 0.027665289118885994, + 0.0011119070695713162, + 0.0012213037116453052, + 0.002145288046449423, + 0.002427060157060623, + 0.0010209613246843219, + 0.0012022563023492694, + 0.0019613897893577814, + 0.0014853589236736298, + 0.0030978014692664146, + 0.0028213125187903643, + 0.003507280955091119, + 0.005987377837300301, + 0.0023060052189975977, + 0.0027729335706681013, + 0.010259836912155151, + 0.005194548051804304, + 0.002454888541251421, + 0.0019724215380847454, + 0.0026269727386534214, + 0.004280788823962212, + 0.0031366082839667797, + 0.0031204461120069027, + 0.00478653609752655, + 0.005195282865315676, + 0.003108444856479764, + 0.0037758038379251957, + 0.004380867816507816, + 0.005281298886984587, + 0.0060412208549678326, + 0.00318276509642601, + 0.006665115710347891, + 0.006093345582485199, + 0.0035613493528217077, + 0.0027455519884824753, + 0.009476029314100742, + 0.005074997898191214, + 0.004938757512718439, + 0.0038226121105253696, + 0.008443835191428661, + 0.00402791565284133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00319126108661294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496998131275177 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.05496998131275177, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496998131275177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46805617213249207, + 0.05931808426976204, + 0.11499054729938507, + 0.033018458634614944, + 0.16439148783683777, + 0.02643236704170704, + 0.04122612625360489, + 0.026641778647899628, + 0.02623799815773964, + 0.039686914533376694 + ], + "credit_map": [ + 0.04983893409371376, + 0.023299233987927437, + 0.06671591848134995, + 0.12695086002349854, + 0.03802456706762314, + 0.026774460449814796, + 0.09205474704504013, + 0.044397465884685516, + 0.009329324588179588, + 0.008033030666410923, + 0.00650362903252244, + 0.010587052442133427, + 0.005050767678767443, + 0.007291566114872694, + 0.0068888128735125065, + 0.005633907858282328, + 0.008851796388626099, + 0.004236660432070494, + 0.04250539839267731, + 0.01669522002339363, + 0.005087694618850946, + 0.0030223061330616474, + 0.028532708063721657, + 0.006058776285499334, + 0.003806166350841522, + 0.0025269282050430775, + 0.008192882873117924, + 0.0032433257438242435, + 0.002717952011153102, + 0.002959037199616432, + 0.006667715031653643, + 0.0029044500552117825, + 0.01707932911813259, + 0.005527540110051632, + 0.06746692210435867, + 0.016599588096141815, + 0.0055752526968717575, + 0.005769752431660891, + 0.032032039016485214, + 0.014341074973344803, + 0.002464827848598361, + 0.003589186118915677, + 0.004274783190339804, + 0.0038701698649674654, + 0.002565324306488037, + 0.002776982029899955, + 0.0033286777324974537, + 0.003562419442459941, + 0.0041123987175524235, + 0.003953790757805109, + 0.004786079283803701, + 0.007045399397611618, + 0.0024735345505177975, + 0.003838073927909136, + 0.009788465686142445, + 0.0052283890545368195, + 0.0025365271139889956, + 0.0031386972405016422, + 0.003707636147737503, + 0.005197080317884684, + 0.002187043195590377, + 0.002039186656475067, + 0.004318812862038612, + 0.0035167979076504707, + 0.0023508628364652395, + 0.002264629118144512, + 0.003096034750342369, + 0.004099322482943535, + 0.0031285190489143133, + 0.0020779301412403584, + 0.004846252501010895, + 0.0043744477443397045, + 0.003815984120592475, + 0.0034437018912285566, + 0.007992987520992756, + 0.005264997016638517, + 0.003574108937755227, + 0.0025900814216583967, + 0.007241788320243359, + 0.0057632713578641415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.05496998131275177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781660065054893 + ], + "opposite_logprob_tokens": [ + -3.5297815799713135 + ], + "logprob_reward": -0.029781660065054893, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781660065054893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2670303285121918, + 0.1303892284631729, + 0.08447682857513428, + 0.052837733179330826, + 0.10610374808311462, + 0.046240005642175674, + 0.10148866474628448, + 0.14233267307281494, + 0.03195204585790634, + 0.037148747593164444 + ], + "credit_map": [ + 0.03710481896996498, + 0.014969093725085258, + 0.028476350009441376, + 0.04612027108669281, + 0.019621362909674644, + 0.016820622608065605, + 0.08138366788625717, + 0.02253413386642933, + 0.008130450733006, + 0.013468838296830654, + 0.015528897754848003, + 0.034231387078762054, + 0.012369127944111824, + 0.008544771932065487, + 0.02583339996635914, + 0.01228234264999628, + 0.008014576509594917, + 0.0070714810863137245, + 0.023986661806702614, + 0.01236626785248518, + 0.005397501401603222, + 0.0045480928383767605, + 0.017427636310458183, + 0.005664619617164135, + 0.005445278249680996, + 0.003945633769035339, + 0.014006161130964756, + 0.005345436278730631, + 0.004978925921022892, + 0.004765124525874853, + 0.009580640122294426, + 0.0047705331817269325, + 0.008280212059617043, + 0.011806356720626354, + 0.022305812686681747, + 0.013544957153499126, + 0.011904839426279068, + 0.009571483358740807, + 0.013907711952924728, + 0.014782379381358624, + 0.007001038640737534, + 0.0035108497831970453, + 0.005533167626708746, + 0.008423097431659698, + 0.0034040557220578194, + 0.005008193664252758, + 0.007119057234376669, + 0.006240544840693474, + 0.009491772390902042, + 0.00796747487038374, + 0.012922207824885845, + 0.016086895018815994, + 0.008563301526010036, + 0.019048815593123436, + 0.013926356099545956, + 0.013481847941875458, + 0.012551520951092243, + 0.014427482150495052, + 0.01596701890230179, + 0.025853456929326057, + 0.013392786495387554, + 0.019024183973670006, + 0.021731490269303322, + 0.019384734332561493, + 0.0033665699884295464, + 0.003242902224883437, + 0.005990747828036547, + 0.004582107067108154, + 0.0036268949043005705, + 0.002879470121115446, + 0.0048684012144804, + 0.003394952742382884, + 0.004284930881112814, + 0.0035901088267564774, + 0.007447727955877781, + 0.004150757100433111, + 0.004745915066450834, + 0.002722640987485647, + 0.006030356977134943, + 0.004176308400928974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.029781660065054893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and gracefully advances, capturing a tranquil courtyard steeped in the elegance of traditional Chinese architecture. Ornate crimson lacquer columns, intricately carved with coiling dragons, support the expansive gray-tiled roof with gracefully curved eaves. In the center of the gray slate court, a lone figure performs a classical dance. She is a woman with long, dark hair arranged in an intricate bun, secured with a delicate jade hairpin, her face a serene, focused expression. She wears a flowing, layered silk Hanfu (Chinese Hanfu) in flowing white, its wide sleeves capturing a gentle air in her movements. With poetic precision, she turns slowly and deliberately, one arm gracefully extended, the translucent silk sleeves unfurling like morning mist. Her movements are a study in controlled fluidity, each step and gesture seamlessly blending into the next, embodying a deep sense of inner peace and discipline. The entire scene is bathed in the soft, diffuse light of a tranquil morning, filtering through a thin mist, casting gentle shadows and highlighting the delicate texture of the silk, creating an atmosphere of profound tranquility and contemplation.", + "seed": 12248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_0/index_5724_seed_12248_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03227020800113678, + "vlm_logprob": -0.03227020800113678, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03227020800113678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007287704385817051 + ], + "opposite_logprob_tokens": [ + -5.507287502288818 + ], + "logprob_reward": -0.007287704385817051, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007287704385817051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15166647732257843, + 0.1171337142586708, + 0.11024598777294159, + 0.06762119382619858, + 0.11305603384971619, + 0.05672811344265938, + 0.04904596507549286, + 0.23154844343662262, + 0.06781511008739471, + 0.035138923674821854 + ], + "credit_map": [ + 0.013255949132144451, + 0.014562356285750866, + 0.019218575209379196, + 0.0230309646576643, + 0.024847112596035004, + 0.019691266119480133, + 0.015905138105154037, + 0.02115512080490589, + 0.010178560391068459, + 0.006932167336344719, + 0.020556462928652763, + 0.015433438122272491, + 0.016807131469249725, + 0.01639322191476822, + 0.01645076274871826, + 0.014381971210241318, + 0.009970540180802345, + 0.011050356552004814, + 0.016250427812337875, + 0.02082986943423748, + 0.011796563863754272, + 0.00919403787702322, + 0.014445784501731396, + 0.016708413138985634, + 0.006325538270175457, + 0.007996714673936367, + 0.0062512862496078014, + 0.012424065731465816, + 0.006938901264220476, + 0.0059694875963032246, + 0.01289890892803669, + 0.008816291578114033, + 0.00890109222382307, + 0.008941089734435081, + 0.00891169998794794, + 0.024021828547120094, + 0.012125109322369099, + 0.009893042035400867, + 0.024459393694996834, + 0.0158027745783329, + 0.005700711160898209, + 0.005647230427712202, + 0.0059913527220487595, + 0.0079524340108037, + 0.0062930467538535595, + 0.006723297759890556, + 0.01096586138010025, + 0.0074541764333844185, + 0.0041801100596785545, + 0.004814385902136564, + 0.006054646801203489, + 0.005874042399227619, + 0.005488989409059286, + 0.00529588432982564, + 0.011091865599155426, + 0.006246039178222418, + 0.009873149916529655, + 0.012545347213745117, + 0.013027479872107506, + 0.07843492180109024, + 0.009636313654482365, + 0.012209853157401085, + 0.022294985130429268, + 0.07352640479803085, + 0.006148975808173418, + 0.008721515536308289, + 0.00914220791310072, + 0.01695813238620758, + 0.005372439511120319, + 0.005725149065256119, + 0.009103505872189999, + 0.006643183063715696, + 0.005619478411972523, + 0.004438561853021383, + 0.005752155091613531, + 0.004506051540374756, + 0.004477309994399548, + 0.0032815320882946253, + 0.0033374910708516836, + 0.0037263433914631605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.007287704385817051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590247586369514 + ], + "opposite_logprob_tokens": [ + -3.895590305328369 + ], + "logprob_reward": -0.020590247586369514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590247586369514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1298949122428894, + 0.05511493980884552, + 0.076811783015728, + 0.03013329952955246, + 0.14347603917121887, + 0.07222244143486023, + 0.06558087468147278, + 0.246964693069458, + 0.045216090977191925, + 0.13458490371704102 + ], + "credit_map": [ + 0.007740646135061979, + 0.008754380978643894, + 0.016898130998015404, + 0.029662834480404854, + 0.02016088180243969, + 0.016076499596238136, + 0.014406789094209671, + 0.01619473472237587, + 0.0048662456683814526, + 0.004618300125002861, + 0.010897266678512096, + 0.011261457577347755, + 0.004536155145615339, + 0.004599813837558031, + 0.005852024536579847, + 0.008483671583235264, + 0.0072030131705105305, + 0.004891914781183004, + 0.013735847547650337, + 0.011410701088607311, + 0.010176031850278378, + 0.006813595071434975, + 0.009227900765836239, + 0.013352773152291775, + 0.0028333894442766905, + 0.0024259083438664675, + 0.0034835950937122107, + 0.004049873445183039, + 0.0034127149265259504, + 0.003205116605386138, + 0.005804623942822218, + 0.0049180747009813786, + 0.007908803410828114, + 0.007514989469200373, + 0.012479116208851337, + 0.03125457838177681, + 0.014703644439578056, + 0.011512565426528454, + 0.0364103838801384, + 0.02169194445014, + 0.005825798027217388, + 0.0061011966317892075, + 0.010360600426793098, + 0.009466949850320816, + 0.009976440109312534, + 0.009603327140212059, + 0.01242868509143591, + 0.008459437638521194, + 0.0059805274941027164, + 0.008216236717998981, + 0.007364342920482159, + 0.010291681624948978, + 0.005567148327827454, + 0.005306138191372156, + 0.013056725263595581, + 0.009798064827919006, + 0.03635771572589874, + 0.018923768773674965, + 0.05157770961523056, + 0.029436787590384483, + 0.019112026318907738, + 0.02433866076171398, + 0.0294857956469059, + 0.037732187658548355, + 0.005577561911195517, + 0.0058133043348789215, + 0.005322631448507309, + 0.009330340661108494, + 0.004135429859161377, + 0.00400971295312047, + 0.004894924815744162, + 0.006132181733846664, + 0.009579873643815517, + 0.013467492535710335, + 0.03231343626976013, + 0.018473686650395393, + 0.014708036556839943, + 0.012301241979002953, + 0.012935320846736431, + 0.020805800333619118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.020590247586369514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779990650713444 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779990650713444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779990650713444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1244918555021286, + 0.039947353303432465, + 0.01630398817360401, + 0.05050919950008392, + 0.2826696038246155, + 0.047901127487421036, + 0.05367999151349068, + 0.14222171902656555, + 0.07487751543521881, + 0.16739757359027863 + ], + "credit_map": [ + 0.00911699328571558, + 0.01153133250772953, + 0.01185731589794159, + 0.028753532096743584, + 0.011690663173794746, + 0.009417658671736717, + 0.02323032356798649, + 0.018894050270318985, + 0.0028686034493148327, + 0.0028605314437299967, + 0.007447976619005203, + 0.005986012984067202, + 0.004229317419230938, + 0.005056960508227348, + 0.005594354122877121, + 0.005903597921133041, + 0.0015061142621561885, + 0.0012262316886335611, + 0.001608650665730238, + 0.0023225913755595684, + 0.0030273515731096268, + 0.002784149022772908, + 0.0018158882157877088, + 0.0020130127668380737, + 0.004069590475410223, + 0.003962482791393995, + 0.004640328697860241, + 0.005835669580847025, + 0.006739696953445673, + 0.0048982310108840466, + 0.012142765335738659, + 0.008220437914133072, + 0.018334651365876198, + 0.015085122548043728, + 0.02662365883588791, + 0.0610094778239727, + 0.030627276748418808, + 0.03137638047337532, + 0.056013330817222595, + 0.043599724769592285, + 0.00331817171536386, + 0.005072837229818106, + 0.007685413584113121, + 0.0073385112918913364, + 0.005329818930476904, + 0.006683016661554575, + 0.006933241151273251, + 0.005540119484066963, + 0.0035668546333909035, + 0.0039095161482691765, + 0.0060388254933059216, + 0.006951568182557821, + 0.005011576693505049, + 0.004760522861033678, + 0.014576048590242863, + 0.008865080773830414, + 0.00738686416298151, + 0.006568131502717733, + 0.009373713284730911, + 0.04545434191823006, + 0.006403114181011915, + 0.007867860607802868, + 0.013331519439816475, + 0.0458361953496933, + 0.006608825642615557, + 0.006933179218322039, + 0.010404885746538639, + 0.01814657635986805, + 0.006060611456632614, + 0.006049791816622019, + 0.011378029361367226, + 0.009295623749494553, + 0.015188036486506462, + 0.01671329140663147, + 0.030650487169623375, + 0.017883971333503723, + 0.025734281167387962, + 0.017176875844597816, + 0.017781775444746017, + 0.02626887522637844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.009779990650713444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10041771829128265 + ], + "opposite_logprob_tokens": [ + -2.3504176139831543 + ], + "logprob_reward": -0.10041771829128265, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10041771829128265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08110357820987701, + 0.05949239060282707, + 0.037937868386507034, + 0.03672969713807106, + 0.09931713342666626, + 0.13920475542545319, + 0.03920562565326691, + 0.3556373715400696, + 0.02680394984781742, + 0.12456757575273514 + ], + "credit_map": [ + 0.008101241663098335, + 0.008150778710842133, + 0.01467149704694748, + 0.013660047203302383, + 0.009709907695651054, + 0.006815018597990274, + 0.007898676209151745, + 0.012096412479877472, + 0.004468460567295551, + 0.0033018740359693766, + 0.010691596195101738, + 0.007781127002090216, + 0.009320475161075592, + 0.0085355956107378, + 0.007641729433089495, + 0.007751531898975372, + 0.0045343912206590176, + 0.0033657075837254524, + 0.005885229911655188, + 0.004662469029426575, + 0.003854466835036874, + 0.0038207804318517447, + 0.006339875981211662, + 0.005474946461617947, + 0.004208706319332123, + 0.0030241552740335464, + 0.003939233720302582, + 0.007345134392380714, + 0.005019140895456076, + 0.003099101362749934, + 0.005458347033709288, + 0.004635876975953579, + 0.008601595647633076, + 0.007489760406315327, + 0.006844986230134964, + 0.020461931824684143, + 0.010982736013829708, + 0.009940197691321373, + 0.021237794309854507, + 0.013758131302893162, + 0.007180747110396624, + 0.007471274118870497, + 0.009274112991988659, + 0.01786334626376629, + 0.016551261767745018, + 0.02326626144349575, + 0.03458581492304802, + 0.023011934012174606, + 0.0029192056972533464, + 0.0033023571595549583, + 0.0035624976735562086, + 0.005266732070595026, + 0.004119163379073143, + 0.006096883676946163, + 0.007735002785921097, + 0.006203785073012114, + 0.015269852243363857, + 0.017265560105443, + 0.017970269545912743, + 0.1103890985250473, + 0.019676407799124718, + 0.021476533263921738, + 0.038332920521497726, + 0.11525674164295197, + 0.0038459300994873047, + 0.002958497265353799, + 0.0032231139484792948, + 0.0031744681764394045, + 0.003950243350118399, + 0.002968672662973404, + 0.00415487214922905, + 0.0025281538255512714, + 0.00738854706287384, + 0.01308909710496664, + 0.020599255338311195, + 0.01624206267297268, + 0.021547900512814522, + 0.015391546301543713, + 0.014542932622134686, + 0.015766238793730736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.10041771829128265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a courtyard featuring traditional Chinese architecture with gray-tiled roofs and curved eaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there ornate crimson lacquer columns with coiling dragon carvings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman wearing a flowing white layered silk Hanfu performing a classical dance in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman have long dark hair styled in an intricate bun secured with a jade hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and gracefully advance through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023275379091501236 + ], + "opposite_logprob_tokens": [ + -3.773275375366211 + ], + "logprob_reward": -0.023275379091501236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023275379091501236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15813647210597992, + 0.03591585531830788, + 0.06709938496351242, + 0.09770648181438446, + 0.06879861652851105, + 0.14115707576274872, + 0.2252543717622757, + 0.1383233219385147, + 0.053272493183612823, + 0.014335903339087963 + ], + "credit_map": [ + 0.009221545420587063, + 0.00974857434630394, + 0.015570072457194328, + 0.03756244108080864, + 0.014996439218521118, + 0.014927838928997517, + 0.026885632425546646, + 0.029223931953310966, + 0.0025073003489524126, + 0.0031625847332179546, + 0.003457818878814578, + 0.008715523406863213, + 0.004915462341159582, + 0.0028894420247524977, + 0.004364507272839546, + 0.0059032198041677475, + 0.004580867011100054, + 0.004415057133883238, + 0.011186669580638409, + 0.011207260191440582, + 0.007599979173392057, + 0.005970820784568787, + 0.011690018698573112, + 0.010448714718222618, + 0.006671159528195858, + 0.006675881799310446, + 0.008625473827123642, + 0.011680693365633488, + 0.014822014607489109, + 0.009381688199937344, + 0.02603304572403431, + 0.013816526159644127, + 0.005816936492919922, + 0.006609957665205002, + 0.008761552162468433, + 0.009915211237967014, + 0.007627745158970356, + 0.012453019618988037, + 0.008843551389873028, + 0.008770649321377277, + 0.011843276210129261, + 0.012269549071788788, + 0.018129313364624977, + 0.012058478780090809, + 0.01679680496454239, + 0.028225690126419067, + 0.02445288747549057, + 0.017381081357598305, + 0.024093763902783394, + 0.02584410458803177, + 0.035533394664525986, + 0.03209897503256798, + 0.019411209970712662, + 0.0196036659181118, + 0.04367881640791893, + 0.02499043568968773, + 0.009223783388733864, + 0.011011941358447075, + 0.0130198635160923, + 0.034122344106435776, + 0.009326805360615253, + 0.008649004623293877, + 0.014361266046762466, + 0.03860831260681152, + 0.008335730992257595, + 0.007332818116992712, + 0.007050591055303812, + 0.008204098790884018, + 0.004707200452685356, + 0.004568778444081545, + 0.0065732537768781185, + 0.006500018294900656, + 0.0012467345222830772, + 0.001260364893823862, + 0.003089283360168338, + 0.0019900884944945574, + 0.002024314599111676, + 0.0012800339609384537, + 0.0018354578642174602, + 0.0016096263425424695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.023275379091501236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_tensor([6948, 6948])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_tensor([6948, 6948])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..27db083e93cca421d086b7ddca0b69e6017edcb5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_tensor([6948, 6948])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, sports-inspired shot, the camera begins at a low, dynamic tracking angle and then slowly ascends to capture an epic clash amidst a turbulent sky. At the center of the frame is a majestic cyan Pegasus, a metallic blue hue. Its feathers, tightly packed like scales forged in steel, gleam with a cool brilliance beneath the dramatic beams of light piercing the clouds. Its most striking features are its mane and tail feathers, not a single color, but a vibrant rainbow spectrum that shimmers with each tumble, gradating from fiery red to deep indigo, radiating a neon inner glow. Engaged in a fierce aerial dogfight, the Pegasus suddenly folds its wings, transforming into a flash of cyan lightning as it dives diagonally downward. Nearing a sea of clouds, it suddenly unfurls its massive wings, harnessing the powerful air currents to launch a near-vertical, gravity-defying ascent. Behind it, its rainbow-colored mane draws long, brilliant trails composed of luminous particles. These trails pause briefly among the thick cumulonimbus clouds, like colorful scars in the sky. As the camera steadily and calmly rises, our field of vision becomes increasingly wider, perfectly capturing every powerful and swift movement of this mythical beast and the grand scale of this tense confrontation, creating a breathtaking, mythical atmosphere of grandeur.", + "seed": 28195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28195_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01908115523401648, + "vlm_logprob": -0.01908115523401648, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.01908115523401648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017997179180383682 + ], + "opposite_logprob_tokens": [ + -4.1429972648620605 + ], + "logprob_reward": -0.017997179180383682, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017997179180383682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12166299670934677, + 0.03252231329679489, + 0.04654783755540848, + 0.07913483679294586, + 0.054849885404109955, + 0.03547121584415436, + 0.11627446860074997, + 0.422396183013916, + 0.056949518620967865, + 0.03419079631567001 + ], + "credit_map": [ + 0.014621156267821789, + 0.013394627720117569, + 0.020769355818629265, + 0.011501847766339779, + 0.015124397352337837, + 0.02270612120628357, + 0.012783201411366463, + 0.01076227705925703, + 0.0037155093159526587, + 0.004076605662703514, + 0.004618276841938496, + 0.004438747651875019, + 0.004832549951970577, + 0.004215474706143141, + 0.003239175770431757, + 0.0033859722316265106, + 0.005176763981580734, + 0.005801803898066282, + 0.0066052996553480625, + 0.00471025425940752, + 0.005103002768009901, + 0.009427044540643692, + 0.005313483066856861, + 0.004410185851156712, + 0.014424501918256283, + 0.009338319301605225, + 0.015004301443696022, + 0.00919406022876501, + 0.008133579045534134, + 0.009653916582465172, + 0.006235689390450716, + 0.007150465156883001, + 0.004948167596012354, + 0.010038148611783981, + 0.007717871107161045, + 0.010348459705710411, + 0.003616988891735673, + 0.006986155640333891, + 0.005767103284597397, + 0.005426985211670399, + 0.003676459426060319, + 0.003983012400567532, + 0.004419288132339716, + 0.0037648985162377357, + 0.004383712541311979, + 0.0055871340446174145, + 0.005418915767222643, + 0.004237791523337364, + 0.010430512949824333, + 0.006759373936802149, + 0.026526693254709244, + 0.011659651063382626, + 0.00963668990880251, + 0.02071690931916237, + 0.011698324233293533, + 0.0188462994992733, + 0.02089480496942997, + 0.0066925943829119205, + 0.18507982790470123, + 0.0687030479311943, + 0.007884280756115913, + 0.010961281135678291, + 0.029655437916517258, + 0.09252488613128662, + 0.0059744445607066154, + 0.005224646534770727, + 0.008488686755299568, + 0.004931184463202953, + 0.005655157379806042, + 0.006488687824457884, + 0.007216754835098982, + 0.012969951145350933, + 0.003966616000980139, + 0.004701027646660805, + 0.004578436724841595, + 0.003830734873190522, + 0.0068504768423736095, + 0.003131631063297391, + 0.003310214728116989, + 0.0038216530811041594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.017997179180383682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001945747178979218 + ], + "opposite_logprob_tokens": [ + -6.251945972442627 + ], + "logprob_reward": -0.001945747178979218, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001945747178979218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10424736887216568, + 0.05716096982359886, + 0.04167316481471062, + 0.040072984993457794, + 0.055910151451826096, + 0.029068194329738617, + 0.141875222325325, + 0.4218955934047699, + 0.058365605771541595, + 0.049730703234672546 + ], + "credit_map": [ + 0.012880699709057808, + 0.010194772854447365, + 0.015424180775880814, + 0.008982406929135323, + 0.010708288289606571, + 0.02538209594786167, + 0.012175983749330044, + 0.008498945273458958, + 0.005776871927082539, + 0.006286397576332092, + 0.006085550412535667, + 0.007282124366611242, + 0.009947169572114944, + 0.00824237521737814, + 0.006296568084508181, + 0.007243913598358631, + 0.00693279504776001, + 0.005164430942386389, + 0.005759758409112692, + 0.004073285963386297, + 0.00428849458694458, + 0.00541740283370018, + 0.004522988572716713, + 0.005514007993042469, + 0.008591779507696629, + 0.004158122930675745, + 0.006401801481842995, + 0.0039552487432956696, + 0.0046729217283427715, + 0.0036954183597117662, + 0.004239245783537626, + 0.004358445294201374, + 0.0058923326432704926, + 0.005911483429372311, + 0.0071904887445271015, + 0.007987956516444683, + 0.003919319249689579, + 0.009451976977288723, + 0.005992313381284475, + 0.009564279578626156, + 0.0022804788313806057, + 0.0029896253254264593, + 0.004190175328403711, + 0.004782017786055803, + 0.0032757120206952095, + 0.0029020006768405437, + 0.005072571337223053, + 0.003575611859560013, + 0.011290635913610458, + 0.006662121042609215, + 0.033244162797927856, + 0.011503702029585838, + 0.010365271009504795, + 0.02446162700653076, + 0.014978460967540741, + 0.029369235038757324, + 0.02220727875828743, + 0.0074325972236692905, + 0.1807333081960678, + 0.06972042471170425, + 0.00975862331688404, + 0.012416848912835121, + 0.023155584931373596, + 0.09647092968225479, + 0.006771591491997242, + 0.006469303276389837, + 0.007631183136254549, + 0.0054475534707307816, + 0.0077040959149599075, + 0.004467630293220282, + 0.0063134003430604935, + 0.013560847379267216, + 0.0057161287404596806, + 0.006480598822236061, + 0.006909591611474752, + 0.004835589323192835, + 0.011453142389655113, + 0.003717412007972598, + 0.005095553584396839, + 0.005522686988115311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.001945747178979218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632889524102211 + ], + "opposite_logprob_tokens": [ + -3.6513288021087646 + ], + "logprob_reward": -0.02632889524102211, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632889524102211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02412211149930954, + 0.017226194962859154, + 0.015489263460040092, + 0.011774934828281403, + 0.019239256158471107, + 0.021462174132466316, + 0.046076398342847824, + 0.769433856010437, + 0.054832715541124344, + 0.02034316398203373 + ], + "credit_map": [ + 0.002780162962153554, + 0.0028302064165472984, + 0.0029877210035920143, + 0.0022973904851824045, + 0.0028881735634058714, + 0.0056444150395691395, + 0.0030255974270403385, + 0.001668442040681839, + 0.0024548748042434454, + 0.002347523346543312, + 0.0044308993965387344, + 0.0017042083200067282, + 0.0010628302115947008, + 0.003048307029530406, + 0.0010419550817459822, + 0.00113559456076473, + 0.0025585342664271593, + 0.0022547796834260225, + 0.002001335145905614, + 0.0015560310566797853, + 0.002041405998170376, + 0.0017627651104703546, + 0.0016004936769604683, + 0.0017139171250164509, + 0.0020924543496221304, + 0.0010682960273697972, + 0.0020204847678542137, + 0.0016831663670018315, + 0.0017077544471248984, + 0.0010746322805061936, + 0.0013452157145366073, + 0.0007829294772818685, + 0.0015724182594567537, + 0.0018524607876315713, + 0.00284397229552269, + 0.003010565647855401, + 0.0017944670980796218, + 0.0026270258240401745, + 0.003127956297248602, + 0.002410389017313719, + 0.001831783214583993, + 0.002050701528787613, + 0.002830972196534276, + 0.0027797638904303312, + 0.0022964528761804104, + 0.0030144532211124897, + 0.0033261270727962255, + 0.003331918967887759, + 0.0036086374893784523, + 0.0031860254239290953, + 0.008827196434140205, + 0.004667873959988356, + 0.0045461393892765045, + 0.007871225476264954, + 0.006910783238708973, + 0.006458512973040342, + 0.03771602362394333, + 0.011457261629402637, + 0.3395960330963135, + 0.129201740026474, + 0.010030626319348812, + 0.019947294145822525, + 0.04746250808238983, + 0.17402231693267822, + 0.00526948319748044, + 0.006724283564835787, + 0.007671603001654148, + 0.005057301837950945, + 0.007930468767881393, + 0.005665726959705353, + 0.009038269519805908, + 0.007475574500858784, + 0.0024213832803070545, + 0.0022685592994093895, + 0.0033775235060602427, + 0.0021722745150327682, + 0.0029100074898451567, + 0.001471425872296095, + 0.002568882657214999, + 0.0031531071290373802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.02632889524102211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011062243953347206 + ], + "opposite_logprob_tokens": [ + -4.511062145233154 + ], + "logprob_reward": -0.011062243953347206, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011062243953347206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03098374791443348, + 0.022680530324578285, + 0.024695098400115967, + 0.019261106848716736, + 0.0536637082695961, + 0.03956565260887146, + 0.04666239395737648, + 0.6185353994369507, + 0.11644459515810013, + 0.027507798746228218 + ], + "credit_map": [ + 0.0030201023910194635, + 0.0039132428355515, + 0.0041189235635101795, + 0.0034469913225620985, + 0.00476933503523469, + 0.0039059713017195463, + 0.0050385482609272, + 0.0027706308756023645, + 0.0033062428701668978, + 0.0021149897947907448, + 0.0062776366248726845, + 0.0024766898714005947, + 0.001144745503552258, + 0.0041319397278130054, + 0.0013268213951960206, + 0.0019014660501852632, + 0.003193765180185437, + 0.0032906648702919483, + 0.0030648591928184032, + 0.0026223936583846807, + 0.0022757998667657375, + 0.004956291522830725, + 0.0027981798630207777, + 0.00249314378015697, + 0.0034036142751574516, + 0.0024127380456775427, + 0.0030583192128688097, + 0.0020803092047572136, + 0.0026006668340414762, + 0.0018085866468027234, + 0.0020257688593119383, + 0.001871103304438293, + 0.006138123106211424, + 0.007062139920890331, + 0.0069769155234098434, + 0.007289526052772999, + 0.0035284103360027075, + 0.006152793765068054, + 0.006594266276806593, + 0.009921533986926079, + 0.0036276716273278, + 0.004655152093619108, + 0.004673622082918882, + 0.005911146756261587, + 0.005000424571335316, + 0.004679572302848101, + 0.005363314412534237, + 0.005654748994857073, + 0.004200974944978952, + 0.003315884154289961, + 0.009142646566033363, + 0.0034513024147599936, + 0.004042610060423613, + 0.008622944355010986, + 0.004357699304819107, + 0.009528334252536297, + 0.0321844108402729, + 0.012366686016321182, + 0.2602788209915161, + 0.10008492320775986, + 0.019967732951045036, + 0.01901775412261486, + 0.032869912683963776, + 0.14176514744758606, + 0.011722014285624027, + 0.01337265595793724, + 0.016917845234274864, + 0.01172968652099371, + 0.014345106668770313, + 0.013373193331062794, + 0.020436519756913185, + 0.014547569677233696, + 0.0032130589243024588, + 0.0038977698422968388, + 0.00388594763353467, + 0.0028378337156027555, + 0.00516124302521348, + 0.00205413019284606, + 0.0026099716778844595, + 0.003847844433039427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.011062243953347206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038071710616350174 + ], + "opposite_logprob_tokens": [ + -3.288071632385254 + ], + "logprob_reward": -0.038071710616350174, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038071710616350174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02515539526939392, + 0.010830704122781754, + 0.010420351289212704, + 0.016928307712078094, + 0.029710588976740837, + 0.06728804111480713, + 0.12154268473386765, + 0.6380255222320557, + 0.06289945542812347, + 0.017198996618390083 + ], + "credit_map": [ + 0.003573688678443432, + 0.0021364863496273756, + 0.00387948309071362, + 0.0021198997274041176, + 0.0022556569892913103, + 0.006830150261521339, + 0.0026416850741952658, + 0.0017183447489514947, + 0.0016992846503853798, + 0.0011068028397858143, + 0.0017312729032710195, + 0.0009833347285166383, + 0.0013493584701791406, + 0.0018984292401000857, + 0.0010833470150828362, + 0.0009788740426301956, + 0.0018139270832762122, + 0.0012991979019716382, + 0.0015505827032029629, + 0.001022642245516181, + 0.0012376324739307165, + 0.0008137760451063514, + 0.0011090057669207454, + 0.0015735867200419307, + 0.003002861747518182, + 0.0021726947743445635, + 0.00375421647913754, + 0.0018071410013362765, + 0.0017219679430127144, + 0.001424722489900887, + 0.0013116727350279689, + 0.0017330287955701351, + 0.003887866623699665, + 0.003040526993572712, + 0.005769402254372835, + 0.00459944037720561, + 0.002233675215393305, + 0.004323235712945461, + 0.002893131924793124, + 0.0029633096419274807, + 0.0057844603434205055, + 0.004399873781949282, + 0.010817576199769974, + 0.006453574635088444, + 0.005577780772000551, + 0.007873660884797573, + 0.01785283163189888, + 0.008528281934559345, + 0.008932994678616524, + 0.0064963181503117085, + 0.036908313632011414, + 0.014185425825417042, + 0.00744345597922802, + 0.015322822146117687, + 0.009160581976175308, + 0.023092759773135185, + 0.031337860971689224, + 0.00959705375134945, + 0.281303733587265, + 0.10774242132902145, + 0.008345205336809158, + 0.01625870354473591, + 0.03894094005227089, + 0.14449958503246307, + 0.006902219727635384, + 0.007542578503489494, + 0.008193613030016422, + 0.005155521910637617, + 0.009264713153243065, + 0.007834821939468384, + 0.009060341864824295, + 0.0089456457644701, + 0.0020285549107939005, + 0.0022691218182444572, + 0.0026374522130936384, + 0.002307942369952798, + 0.003084000898525119, + 0.001407134928740561, + 0.001823514117859304, + 0.0016412733821198344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.038071710616350174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, sports-inspired shot, the camera begins at a low, dynamic tracking angle and then slowly ascends to capture an epic clash amidst a turbulent sky. At the center of the frame is a majestic cyan Pegasus, a metallic blue hue. Its feathers, tightly packed like scales forged in steel, gleam with a cool brilliance beneath the dramatic beams of light piercing the clouds. Its most striking features are its mane and tail feathers, not a single color, but a vibrant rainbow spectrum that shimmers with each tumble, gradating from fiery red to deep indigo, radiating a neon inner glow. Engaged in a fierce aerial dogfight, the Pegasus suddenly folds its wings, transforming into a flash of cyan lightning as it dives diagonally downward. Nearing a sea of clouds, it suddenly unfurls its massive wings, harnessing the powerful air currents to launch a near-vertical, gravity-defying ascent. Behind it, its rainbow-colored mane draws long, brilliant trails composed of luminous particles. These trails pause briefly among the thick cumulonimbus clouds, like colorful scars in the sky. As the camera steadily and calmly rises, our field of vision becomes increasingly wider, perfectly capturing every powerful and swift movement of this mythical beast and the grand scale of this tense confrontation, creating a breathtaking, mythical atmosphere of grandeur.", + "seed": 28196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28196_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016776741482317448, + "vlm_logprob": -0.016776741482317448, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.016776741482317448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010350477881729603 + ], + "opposite_logprob_tokens": [ + -4.760350704193115 + ], + "logprob_reward": -0.010350477881729603, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010350477881729603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05421724170446396, + 0.05171344056725502, + 0.031603433191776276, + 0.5634637475013733, + 0.07223064452409744, + 0.03709554672241211, + 0.05115969106554985, + 0.03180904686450958, + 0.04260420799255371, + 0.06410287320613861 + ], + "credit_map": [ + 0.005768434144556522, + 0.006853876169770956, + 0.0038601201958954334, + 0.005611961241811514, + 0.005882080178707838, + 0.009776096791028976, + 0.006455665919929743, + 0.010009022429585457, + 0.005086161196231842, + 0.0058141411282122135, + 0.010181549936532974, + 0.006136737298220396, + 0.0069123380817472935, + 0.005696138367056847, + 0.005129569210112095, + 0.006756816990673542, + 0.0036705967504531145, + 0.0023810493294149637, + 0.004021251108497381, + 0.0036650891415774822, + 0.004171785432845354, + 0.0046438658609986305, + 0.004024931229650974, + 0.0050248680636286736, + 0.1487828493118286, + 0.03898758813738823, + 0.0419854000210762, + 0.040967728942632675, + 0.024266136810183525, + 0.20261962711811066, + 0.029835980385541916, + 0.036018550395965576, + 0.006732262205332518, + 0.011126885190606117, + 0.00873432494699955, + 0.006202496122568846, + 0.010077029466629028, + 0.011032010428607464, + 0.007431636564433575, + 0.01089401263743639, + 0.004951991140842438, + 0.005760303232818842, + 0.005693899467587471, + 0.003883536672219634, + 0.005026553757488728, + 0.0028826468624174595, + 0.004232365638017654, + 0.004664259031414986, + 0.0069317505694925785, + 0.0063518197275698185, + 0.0056679570116102695, + 0.006502613425254822, + 0.004871645011007786, + 0.00524760689586401, + 0.006183729972690344, + 0.009402577765285969, + 0.0024098448920994997, + 0.005222912412136793, + 0.0035197448451071978, + 0.004981313366442919, + 0.005469569470733404, + 0.0037018039729446173, + 0.00372211798094213, + 0.0027817499358206987, + 0.0037717854138463736, + 0.005842224694788456, + 0.006015053018927574, + 0.005504610016942024, + 0.006725691724568605, + 0.004989737179130316, + 0.004744779784232378, + 0.005010333377867937, + 0.004370637238025665, + 0.01020065601915121, + 0.0056905560195446014, + 0.0078019495122134686, + 0.01474432647228241, + 0.008706492371857166, + 0.005405695177614689, + 0.007182571571320295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.010350477881729603 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254387855529785 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254387855529785, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254387855529785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0831754207611084, + 0.044325314462184906, + 0.11607211828231812, + 0.4527176320552826, + 0.07152675837278366, + 0.050986725836992264, + 0.04264025390148163, + 0.044037871062755585, + 0.05090758204460144, + 0.0436103381216526 + ], + "credit_map": [ + 0.007615251932293177, + 0.009863210842013359, + 0.009721904061734676, + 0.007776289246976376, + 0.008818312548100948, + 0.014534582383930683, + 0.008671111427247524, + 0.01617475226521492, + 0.005087162833660841, + 0.007113327272236347, + 0.007741335313767195, + 0.00458574341610074, + 0.006223954260349274, + 0.005259048193693161, + 0.00440549198538065, + 0.003909253515303135, + 0.010501776821911335, + 0.01320144347846508, + 0.014072696678340435, + 0.016121774911880493, + 0.013443950563669205, + 0.01576700061559677, + 0.014099661260843277, + 0.018863815814256668, + 0.1191985160112381, + 0.031998854130506516, + 0.03166321665048599, + 0.030671847984194756, + 0.018816426396369934, + 0.15720534324645996, + 0.02717067487537861, + 0.03599275276064873, + 0.008974414318799973, + 0.006480989046394825, + 0.014751414768397808, + 0.007676081266254187, + 0.008985298685729504, + 0.010674024000763893, + 0.00575935747474432, + 0.00822517555207014, + 0.006016034632921219, + 0.00690907659009099, + 0.008490176871418953, + 0.0061326525174081326, + 0.005980972666293383, + 0.004954777657985687, + 0.005333114881068468, + 0.007169922813773155, + 0.005447858944535255, + 0.005528654437512159, + 0.00495436554774642, + 0.0049156625755131245, + 0.0058072651736438274, + 0.0047649298794567585, + 0.005263322498649359, + 0.005958189256489277, + 0.0046351137571036816, + 0.006467617116868496, + 0.004685719031840563, + 0.005351367872208357, + 0.007483681198209524, + 0.006033400073647499, + 0.005034233443439007, + 0.004346739966422319, + 0.004431216977536678, + 0.005745113827288151, + 0.004753408953547478, + 0.00785196665674448, + 0.006921934895217419, + 0.006895828060805798, + 0.004996269941329956, + 0.009311842732131481, + 0.0027423046994954348, + 0.0062229083850979805, + 0.0051635196432471275, + 0.003746492089703679, + 0.009452639147639275, + 0.006521902978420258, + 0.003249744651839137, + 0.006510828156024218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.005254387855529785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066370643675327 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066370643675327, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066370643675327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07058225572109222, + 0.08865112066268921, + 0.03705614060163498, + 0.2501317858695984, + 0.08828894793987274, + 0.11163980513811111, + 0.06848607957363129, + 0.07216113805770874, + 0.09956172853708267, + 0.11344094574451447 + ], + "credit_map": [ + 0.006502422038465738, + 0.008417476899921894, + 0.005891628563404083, + 0.006396380253136158, + 0.009748931042850018, + 0.01471792533993721, + 0.008707858622074127, + 0.010199635289609432, + 0.011326261796057224, + 0.01045211125165224, + 0.016777269542217255, + 0.010697834193706512, + 0.008941911160945892, + 0.00860289391130209, + 0.011580069549381733, + 0.010272776708006859, + 0.004536827094852924, + 0.0033988317009061575, + 0.005654239561408758, + 0.003588781924918294, + 0.006121515762060881, + 0.004241243004798889, + 0.004130838438868523, + 0.005383865907788277, + 0.06006920337677002, + 0.016788719221949577, + 0.023447250947356224, + 0.019440440461039543, + 0.010619666427373886, + 0.07845214009284973, + 0.020794402807950974, + 0.020519981160759926, + 0.009424940682947636, + 0.012250608764588833, + 0.010661840438842773, + 0.00869100820273161, + 0.01123188715428114, + 0.012110704556107521, + 0.006873708218336105, + 0.01704426109790802, + 0.010673743672668934, + 0.014632550068199635, + 0.01529074739664793, + 0.01145236101001501, + 0.015633786097168922, + 0.008883817121386528, + 0.017109589651226997, + 0.017963219434022903, + 0.00933122355490923, + 0.005777362268418074, + 0.007456197403371334, + 0.008809782564640045, + 0.007349417544901371, + 0.007878671400249004, + 0.010390430688858032, + 0.011492997407913208, + 0.006373930256813765, + 0.013199113309383392, + 0.008752907626330853, + 0.009268611669540405, + 0.01098385639488697, + 0.008957469835877419, + 0.0060784281231462955, + 0.008546821773052216, + 0.008443912491202354, + 0.013233068399131298, + 0.0077389907091856, + 0.012286476790904999, + 0.01584474928677082, + 0.014823632314801216, + 0.009065207093954086, + 0.01812569983303547, + 0.009549240581691265, + 0.015005014836788177, + 0.014030978083610535, + 0.008711010217666626, + 0.02215178869664669, + 0.017254149541258812, + 0.008110803551971912, + 0.018627967685461044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.011066370643675327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014179903082549572 + ], + "opposite_logprob_tokens": [ + -4.264179706573486 + ], + "logprob_reward": -0.014179903082549572, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014179903082549572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05943857878446579, + 0.04348378628492355, + 0.030237408354878426, + 0.40356430411338806, + 0.11963900923728943, + 0.04211495444178581, + 0.02923417091369629, + 0.1104021668434143, + 0.07213503122329712, + 0.08975065499544144 + ], + "credit_map": [ + 0.0052225012332201, + 0.007332317065447569, + 0.00587590504437685, + 0.005936532747000456, + 0.008091963827610016, + 0.012045549228787422, + 0.007054510526359081, + 0.007879291661083698, + 0.005365785211324692, + 0.004556847270578146, + 0.008476468734443188, + 0.004319552797824144, + 0.005494688171893358, + 0.004526939708739519, + 0.004939358681440353, + 0.005804144311696291, + 0.003870984073728323, + 0.002685354556888342, + 0.004240819253027439, + 0.002784654265269637, + 0.002657586243003607, + 0.005259902682155371, + 0.002800493035465479, + 0.005937611684203148, + 0.11149834096431732, + 0.026768993586301804, + 0.027880804613232613, + 0.026219308376312256, + 0.01337334606796503, + 0.1499003767967224, + 0.023535002022981644, + 0.02438809722661972, + 0.013812794350087643, + 0.011734710074961185, + 0.021604949608445168, + 0.012365316972136497, + 0.011086643673479557, + 0.015394788235425949, + 0.011423365212976933, + 0.0222164336591959, + 0.0043861460871994495, + 0.00592180946841836, + 0.00740435253828764, + 0.005364477168768644, + 0.004640466533601284, + 0.0034599292557686567, + 0.004365607630461454, + 0.006572159472852945, + 0.003183856373652816, + 0.0030871641356498003, + 0.0037444669287651777, + 0.004062345717102289, + 0.004454909823834896, + 0.002997639821842313, + 0.004569750279188156, + 0.003134035272523761, + 0.012010494247078896, + 0.023260608315467834, + 0.012498780153691769, + 0.01290606427937746, + 0.016415392979979515, + 0.013532528653740883, + 0.010587959550321102, + 0.00919033121317625, + 0.006415341515094042, + 0.009146354161202908, + 0.006399204954504967, + 0.010718266479671001, + 0.010565021075308323, + 0.01022319309413433, + 0.0072775594890117645, + 0.011390085332095623, + 0.005643663462251425, + 0.013915127143263817, + 0.00612667016685009, + 0.007363422773778439, + 0.02409428171813488, + 0.010612089186906815, + 0.009865939617156982, + 0.012129459530115128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.014179903082549572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303256794810295 + ], + "opposite_logprob_tokens": [ + -3.168032646179199 + ], + "logprob_reward": -0.04303256794810295, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303256794810295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06802031397819519, + 0.11457527428865433, + 0.03730686008930206, + 0.35463836789131165, + 0.1497613936662674, + 0.05918252095580101, + 0.036373570561409, + 0.029040386900305748, + 0.09491024911403656, + 0.056191034615039825 + ], + "credit_map": [ + 0.0063271354883909225, + 0.008008583448827267, + 0.005780360195785761, + 0.0068387221544981, + 0.009367413818836212, + 0.011267290450632572, + 0.00873584859073162, + 0.011694956570863724, + 0.014733975753188133, + 0.013827160000801086, + 0.01968386583030224, + 0.013013327494263649, + 0.010583871975541115, + 0.013325556181371212, + 0.014118603430688381, + 0.015288922935724258, + 0.003621730487793684, + 0.00252345553599298, + 0.005548132583498955, + 0.0033774427138268948, + 0.0033699041232466698, + 0.00690698716789484, + 0.003572455607354641, + 0.008386754430830479, + 0.09037236869335175, + 0.022460192441940308, + 0.028858546167612076, + 0.024578332901000977, + 0.01155478972941637, + 0.11678259819746017, + 0.025277690961956978, + 0.03475385159254074, + 0.014501761645078659, + 0.024279523640871048, + 0.02037050388753414, + 0.014134438708424568, + 0.021995434537529945, + 0.022384170442819595, + 0.010294253937900066, + 0.021801304072141647, + 0.00508540403097868, + 0.008147846907377243, + 0.011272663250565529, + 0.00833845604211092, + 0.006953487638384104, + 0.005219200160354376, + 0.005826614797115326, + 0.008338847197592258, + 0.00565401092171669, + 0.0038433794397860765, + 0.004863006062805653, + 0.005179726518690586, + 0.004351911135017872, + 0.0035533765330910683, + 0.004200458526611328, + 0.004727700725197792, + 0.002847085241228342, + 0.0029937170911580324, + 0.0035239383578300476, + 0.0053825885988771915, + 0.004232985433191061, + 0.0028603901155292988, + 0.0038176835514605045, + 0.00338199851103127, + 0.0073786042630672455, + 0.013358431868255138, + 0.01613747701048851, + 0.015713661909103394, + 0.012282395735383034, + 0.009786861948668957, + 0.010905110277235508, + 0.00934770330786705, + 0.005121270194649696, + 0.00923248752951622, + 0.006394696421921253, + 0.006156726740300655, + 0.012382978573441505, + 0.005317352246493101, + 0.004848864860832691, + 0.006736658979207277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.04303256794810295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, sports-inspired shot, the camera begins at a low, dynamic tracking angle and then slowly ascends to capture an epic clash amidst a turbulent sky. At the center of the frame is a majestic cyan Pegasus, a metallic blue hue. Its feathers, tightly packed like scales forged in steel, gleam with a cool brilliance beneath the dramatic beams of light piercing the clouds. Its most striking features are its mane and tail feathers, not a single color, but a vibrant rainbow spectrum that shimmers with each tumble, gradating from fiery red to deep indigo, radiating a neon inner glow. Engaged in a fierce aerial dogfight, the Pegasus suddenly folds its wings, transforming into a flash of cyan lightning as it dives diagonally downward. Nearing a sea of clouds, it suddenly unfurls its massive wings, harnessing the powerful air currents to launch a near-vertical, gravity-defying ascent. Behind it, its rainbow-colored mane draws long, brilliant trails composed of luminous particles. These trails pause briefly among the thick cumulonimbus clouds, like colorful scars in the sky. As the camera steadily and calmly rises, our field of vision becomes increasingly wider, perfectly capturing every powerful and swift movement of this mythical beast and the grand scale of this tense confrontation, creating a breathtaking, mythical atmosphere of grandeur.", + "seed": 28197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28197_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01562529061920941, + "vlm_logprob": -0.01562529061920941, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01562529061920941, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012791253626346588 + ], + "opposite_logprob_tokens": [ + -4.512791156768799 + ], + "logprob_reward": -0.012791253626346588, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012791253626346588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12030304968357086, + 0.08050210028886795, + 0.05045458301901817, + 0.21372124552726746, + 0.022711409255862236, + 0.08893254399299622, + 0.060449082404375076, + 0.16671183705329895, + 0.13902340829372406, + 0.05719076842069626 + ], + "credit_map": [ + 0.013297991827130318, + 0.014142073690891266, + 0.011073348112404346, + 0.01763041317462921, + 0.02034148946404457, + 0.016598302870988846, + 0.011169628240168095, + 0.01604979857802391, + 0.00939100980758667, + 0.011462343856692314, + 0.00580181460827589, + 0.005999547895044088, + 0.010611403733491898, + 0.011048861779272556, + 0.007893256843090057, + 0.01829386129975319, + 0.006406534928828478, + 0.005530679132789373, + 0.00860073696821928, + 0.003603153396397829, + 0.0032337764278054237, + 0.008352929726243019, + 0.005442209076136351, + 0.009284565225243568, + 0.012048670090734959, + 0.07925888895988464, + 0.0063107130117714405, + 0.005892607383430004, + 0.009127202443778515, + 0.07741247862577438, + 0.009932895191013813, + 0.013737787492573261, + 0.0023910864256322384, + 0.0023476642090827227, + 0.0030988273210823536, + 0.0028531658463180065, + 0.004516639281064272, + 0.0020968662574887276, + 0.0026474047917872667, + 0.002759755589067936, + 0.0055694920010864735, + 0.013520698994398117, + 0.011994848027825356, + 0.011076843366026878, + 0.010393169708549976, + 0.010049830190837383, + 0.008537311106920242, + 0.017790351063013077, + 0.006220769137144089, + 0.007389702834188938, + 0.005877506919205189, + 0.010701571591198444, + 0.009168132208287716, + 0.0068223681300878525, + 0.006422495469450951, + 0.007846537046134472, + 0.01219285186380148, + 0.012699485756456852, + 0.04427960515022278, + 0.007698477245867252, + 0.012945052236318588, + 0.015183882787823677, + 0.008741605095565319, + 0.052970871329307556, + 0.01305041741579771, + 0.015642723068594933, + 0.01440465822815895, + 0.013133092783391476, + 0.018629010766744614, + 0.028947286307811737, + 0.010555315762758255, + 0.024660898372530937, + 0.006953097879886627, + 0.008324573747813702, + 0.0070179663598537445, + 0.004528735764324665, + 0.009787231683731079, + 0.007517206482589245, + 0.005949780810624361, + 0.007112175226211548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.012791253626346588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09501595050096512, + 0.21168488264083862, + 0.09453458338975906, + 0.0775347575545311, + 0.1051713153719902, + 0.11168555915355682, + 0.04812874272465706, + 0.1292000114917755, + 0.057381581515073776, + 0.06966257095336914 + ], + "credit_map": [ + 0.01226852647960186, + 0.009221655316650867, + 0.00991110596805811, + 0.012357041239738464, + 0.016446327790617943, + 0.013973366469144821, + 0.007781278807669878, + 0.013056647032499313, + 0.020776871591806412, + 0.049106892198324203, + 0.016188593581318855, + 0.02011045627295971, + 0.03226076811552048, + 0.02257429249584675, + 0.01953789032995701, + 0.03112911619246006, + 0.01424781046807766, + 0.012791846878826618, + 0.020994914695620537, + 0.008692478761076927, + 0.0052338819950819016, + 0.013889596797525883, + 0.007431709207594395, + 0.011252342723309994, + 0.007827101275324821, + 0.013928351923823357, + 0.008204899728298187, + 0.005081465933471918, + 0.005527522414922714, + 0.016126446425914764, + 0.011246639303863049, + 0.009592329151928425, + 0.00899568758904934, + 0.00931647326797247, + 0.014665825292468071, + 0.01718628592789173, + 0.016168540343642235, + 0.010204590857028961, + 0.01481211744248867, + 0.01382179744541645, + 0.005719711072742939, + 0.008642151951789856, + 0.01906479336321354, + 0.01816732808947563, + 0.012777617201209068, + 0.011263052001595497, + 0.01161507423967123, + 0.024435829371213913, + 0.004269530531018972, + 0.00725827319547534, + 0.004429875873029232, + 0.008488839492201805, + 0.0067392499186098576, + 0.00506177544593811, + 0.0050656418316066265, + 0.006815553642809391, + 0.009352256543934345, + 0.015261461958289146, + 0.023409662768244743, + 0.00716412765905261, + 0.01561774592846632, + 0.01731964386999607, + 0.0098399193957448, + 0.031235188245773315, + 0.006966566666960716, + 0.006445591803640127, + 0.007177483290433884, + 0.004747934173792601, + 0.007829293608665466, + 0.00940888561308384, + 0.004883521236479282, + 0.00992230512201786, + 0.007535803131759167, + 0.007810525130480528, + 0.00934826023876667, + 0.0062293363735079765, + 0.016193101182579994, + 0.009068437851965427, + 0.006010731682181358, + 0.0074663772247731686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533823028206825 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533823028206825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533823028206825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060748420655727386, + 0.08160160481929779, + 0.06109600141644478, + 0.21043211221694946, + 0.04232219234108925, + 0.07418406754732132, + 0.038038503378629684, + 0.15409915149211884, + 0.21642033755779266, + 0.061057645827531815 + ], + "credit_map": [ + 0.006445488426834345, + 0.005288754589855671, + 0.005356105975806713, + 0.006070881150662899, + 0.008246914483606815, + 0.01013835147023201, + 0.004154588095843792, + 0.015047329477965832, + 0.00694082397967577, + 0.008913978934288025, + 0.00530166644603014, + 0.00551543990150094, + 0.01613682694733143, + 0.012675507925450802, + 0.008499670773744583, + 0.017617683857679367, + 0.008400725200772285, + 0.00867892149835825, + 0.008471839129924774, + 0.004508992191404104, + 0.00330534391105175, + 0.01042409148067236, + 0.005126804579049349, + 0.012179279699921608, + 0.01236114650964737, + 0.07744564861059189, + 0.005117638502269983, + 0.0054248315282166, + 0.00827574823051691, + 0.0834430530667305, + 0.006561442278325558, + 0.01180257834494114, + 0.0053570885211229324, + 0.0035486642736941576, + 0.007266792934387922, + 0.006012612022459507, + 0.005006676539778709, + 0.004422803409397602, + 0.004734898917376995, + 0.005972649902105331, + 0.005008232779800892, + 0.010689262300729752, + 0.010536544024944305, + 0.00802990235388279, + 0.010139117948710918, + 0.00921155046671629, + 0.008742314763367176, + 0.01182713732123375, + 0.004056642763316631, + 0.004530578386038542, + 0.003838967764750123, + 0.0061958348378539085, + 0.00507960794493556, + 0.004501352086663246, + 0.0036204420030117035, + 0.00621507503092289, + 0.010495356284081936, + 0.008934015408158302, + 0.04505731165409088, + 0.006281845271587372, + 0.009028204716742039, + 0.015020505525171757, + 0.007518918253481388, + 0.05176296457648277, + 0.02048097923398018, + 0.022112421691417694, + 0.029116902500391006, + 0.013967152684926987, + 0.02123759314417839, + 0.041108742356300354, + 0.016112279146909714, + 0.05228424817323685, + 0.007943162694573402, + 0.008566202595829964, + 0.008984532207250595, + 0.005650658160448074, + 0.00811593234539032, + 0.007008305750787258, + 0.005915890913456678, + 0.008872956968843937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.012533823028206825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012526288628578186 + ], + "opposite_logprob_tokens": [ + -4.387526512145996 + ], + "logprob_reward": -0.012526288628578186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012526288628578186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05905032157897949, + 0.09879318624734879, + 0.041678380221128464, + 0.18384133279323578, + 0.043390508741140366, + 0.09414046257734299, + 0.09151569753885269, + 0.13016942143440247, + 0.10689324885606766, + 0.1505274474620819 + ], + "credit_map": [ + 0.009337655268609524, + 0.005385959055274725, + 0.006653417833149433, + 0.005166045390069485, + 0.009659798815846443, + 0.009545529261231422, + 0.004344393964856863, + 0.008957522921264172, + 0.009067176841199398, + 0.02109704539179802, + 0.006351288408041, + 0.008138955570757389, + 0.016217131167650223, + 0.012383535504341125, + 0.009725159965455532, + 0.015812896192073822, + 0.004335483536124229, + 0.004964341409504414, + 0.010052962228655815, + 0.0028109948616474867, + 0.0023385118693113327, + 0.006920250598341227, + 0.004354518838226795, + 0.005901316646486521, + 0.015029079280793667, + 0.05557464063167572, + 0.005691914353519678, + 0.00717879319563508, + 0.006414511706680059, + 0.06958659738302231, + 0.01404552347958088, + 0.010320269502699375, + 0.004936338402330875, + 0.005301501136273146, + 0.005861811805516481, + 0.005502194166183472, + 0.005371830426156521, + 0.006217624060809612, + 0.0051069376058876514, + 0.005092271603643894, + 0.006935719400644302, + 0.009467204101383686, + 0.014053981751203537, + 0.010550782084465027, + 0.012647663243114948, + 0.011546807363629341, + 0.011226143687963486, + 0.01771216280758381, + 0.0073215593583881855, + 0.008855591528117657, + 0.010266026481986046, + 0.01494503766298294, + 0.008653238415718079, + 0.012318155728280544, + 0.010994638316333294, + 0.018161451444029808, + 0.011281969025731087, + 0.017726227641105652, + 0.012438341043889523, + 0.01622673124074936, + 0.019734900444746017, + 0.016749540343880653, + 0.013909587636590004, + 0.02210213616490364, + 0.011055945418775082, + 0.012299859896302223, + 0.013289329595863819, + 0.008209100924432278, + 0.01817290671169758, + 0.01637168787419796, + 0.008664059452712536, + 0.01883035898208618, + 0.02285218983888626, + 0.025483742356300354, + 0.018746117129921913, + 0.013470103964209557, + 0.018464000895619392, + 0.017857590690255165, + 0.012737700715661049, + 0.020915990695357323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.012526288628578186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380728580057621 + ], + "opposite_logprob_tokens": [ + -3.2880728244781494 + ], + "logprob_reward": -0.0380728580057621, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380728580057621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03798164054751396, + 0.05307772383093834, + 0.028499284759163857, + 0.3002844750881195, + 0.014323646202683449, + 0.0518190823495388, + 0.06414005160331726, + 0.0812552347779274, + 0.29398924112319946, + 0.07462961226701736 + ], + "credit_map": [ + 0.00397515669465065, + 0.003125167218968272, + 0.002875617239624262, + 0.0021839363034814596, + 0.007143207360059023, + 0.00689993891865015, + 0.0027229429688304663, + 0.00905567780137062, + 0.004681820515543222, + 0.005579531192779541, + 0.0038251522928476334, + 0.004576584789901972, + 0.01129209529608488, + 0.007749884855002165, + 0.00513789989054203, + 0.01023476105183363, + 0.003215787000954151, + 0.0031962303910404444, + 0.0041574230417609215, + 0.0018774783238768578, + 0.001750783994793892, + 0.005337847396731377, + 0.0027713149320334196, + 0.006192419212311506, + 0.020096048712730408, + 0.11710303276777267, + 0.006940959952771664, + 0.0072851949371397495, + 0.010169217362999916, + 0.11086110770702362, + 0.010748635977506638, + 0.017080295830965042, + 0.0013878874015063047, + 0.0016535179456695914, + 0.0017713775159791112, + 0.0017059631645679474, + 0.0022003985941410065, + 0.0021227276884019375, + 0.0018048692727461457, + 0.0016769065987318754, + 0.004400900565087795, + 0.007515395991504192, + 0.007932080887258053, + 0.005442303139716387, + 0.00579009810462594, + 0.005272040609270334, + 0.005613305140286684, + 0.009852960705757141, + 0.004820677917450666, + 0.0045015569776296616, + 0.0075361113995313644, + 0.009655892848968506, + 0.006455427035689354, + 0.008122490718960762, + 0.00869352463632822, + 0.014354373328387737, + 0.006583280861377716, + 0.011375492438673973, + 0.013166040182113647, + 0.004600478336215019, + 0.010464215651154518, + 0.012958052568137646, + 0.00854866299778223, + 0.013559022918343544, + 0.025528235360980034, + 0.030292723327875137, + 0.03832051903009415, + 0.018128810450434685, + 0.030616436153650284, + 0.06029710918664932, + 0.020411184057593346, + 0.07039423286914825, + 0.009382939897477627, + 0.011059176176786423, + 0.011136680841445923, + 0.006211535539478064, + 0.008344084024429321, + 0.008952935226261616, + 0.007244896609336138, + 0.01229737140238285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.0380728580057621 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, sports-inspired shot, the camera begins at a low, dynamic tracking angle and then slowly ascends to capture an epic clash amidst a turbulent sky. At the center of the frame is a majestic cyan Pegasus, a metallic blue hue. Its feathers, tightly packed like scales forged in steel, gleam with a cool brilliance beneath the dramatic beams of light piercing the clouds. Its most striking features are its mane and tail feathers, not a single color, but a vibrant rainbow spectrum that shimmers with each tumble, gradating from fiery red to deep indigo, radiating a neon inner glow. Engaged in a fierce aerial dogfight, the Pegasus suddenly folds its wings, transforming into a flash of cyan lightning as it dives diagonally downward. Nearing a sea of clouds, it suddenly unfurls its massive wings, harnessing the powerful air currents to launch a near-vertical, gravity-defying ascent. Behind it, its rainbow-colored mane draws long, brilliant trails composed of luminous particles. These trails pause briefly among the thick cumulonimbus clouds, like colorful scars in the sky. As the camera steadily and calmly rises, our field of vision becomes increasingly wider, perfectly capturing every powerful and swift movement of this mythical beast and the grand scale of this tense confrontation, creating a breathtaking, mythical atmosphere of grandeur.", + "seed": 28198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_1/index_6948_seed_28198_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01699129370972514, + "vlm_logprob": -0.01699129370972514, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01699129370972514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005820113234221935 + ], + "opposite_logprob_tokens": [ + -5.380820274353027 + ], + "logprob_reward": -0.005820113234221935, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005820113234221935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13655434548854828, + 0.029885556548833847, + 0.16323348879814148, + 0.030489923432469368, + 0.22655725479125977, + 0.05158080533146858, + 0.09027921408414841, + 0.15027004480361938, + 0.07282798737287521, + 0.04832141101360321 + ], + "credit_map": [ + 0.015023344196379185, + 0.010352889075875282, + 0.01880350336432457, + 0.02223566174507141, + 0.017442073673009872, + 0.02037823386490345, + 0.015355680137872696, + 0.01696295104920864, + 0.002817924367263913, + 0.003221038496121764, + 0.0035533467307686806, + 0.004766786005347967, + 0.005315064452588558, + 0.002262615133076906, + 0.003729121759533882, + 0.004219655878841877, + 0.017533376812934875, + 0.03405505791306496, + 0.020321674644947052, + 0.019957100972533226, + 0.020806413143873215, + 0.019647974520921707, + 0.011728081852197647, + 0.01918378844857216, + 0.004712352529168129, + 0.003353831358253956, + 0.0037846274208277464, + 0.0029240711592137814, + 0.005313005298376083, + 0.0033691280987113714, + 0.0029923678375780582, + 0.004040537867695093, + 0.017731323838233948, + 0.022514043375849724, + 0.031695928424596786, + 0.026385532692074776, + 0.026819586753845215, + 0.030189067125320435, + 0.02131185494363308, + 0.04990990832448006, + 0.003826855681836605, + 0.0027711207512766123, + 0.0058131031692028046, + 0.0062622977420687675, + 0.0066055841743946075, + 0.006561637856066227, + 0.009435757994651794, + 0.010304445400834084, + 0.00952767115086317, + 0.005136412568390369, + 0.010870863683521748, + 0.009260960854589939, + 0.005889778956770897, + 0.012938237749040127, + 0.02795313112437725, + 0.008702150546014309, + 0.017925115302205086, + 0.011479153297841549, + 0.014648936688899994, + 0.03978325054049492, + 0.008696004748344421, + 0.016153642907738686, + 0.030286774039268494, + 0.011297158896923065, + 0.00584561750292778, + 0.011375755071640015, + 0.006062936969101429, + 0.010891348123550415, + 0.012037929147481918, + 0.007827620953321457, + 0.011389586143195629, + 0.007397182285785675, + 0.005094622727483511, + 0.006254116538912058, + 0.009065727703273296, + 0.005904576275497675, + 0.006137070711702108, + 0.0056936899200081825, + 0.006208615843206644, + 0.003962985239923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.005820113234221935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525687774643302 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525687774643302, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525687774643302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09171970188617706, + 0.06526122242212296, + 0.12326538562774658, + 0.22088579833507538, + 0.050904612988233566, + 0.04206657037138939, + 0.17025724053382874, + 0.059298075735569, + 0.12865842878818512, + 0.047683026641607285 + ], + "credit_map": [ + 0.012185444124042988, + 0.009716051630675793, + 0.008800354786217213, + 0.012170101515948772, + 0.009585386142134666, + 0.015208165161311626, + 0.008997127413749695, + 0.015057068318128586, + 0.006105510983616114, + 0.00871195923537016, + 0.005545750260353088, + 0.011109917424619198, + 0.009015774354338646, + 0.0065484242513775826, + 0.008143968880176544, + 0.010079912841320038, + 0.011731117963790894, + 0.01921665295958519, + 0.014427946880459785, + 0.016951799392700195, + 0.018738767132163048, + 0.014644862152636051, + 0.013212885707616806, + 0.014341339468955994, + 0.0184937696903944, + 0.02039157971739769, + 0.046377331018447876, + 0.028902210295200348, + 0.025588829070329666, + 0.029402241110801697, + 0.01759936846792698, + 0.03413046523928642, + 0.004242390859872103, + 0.004278151784092188, + 0.009109528735280037, + 0.005847674794495106, + 0.007909899577498436, + 0.007099991198629141, + 0.004264663439244032, + 0.008152308873832226, + 0.003119890345260501, + 0.002904986497014761, + 0.0051308912225067616, + 0.00479081878438592, + 0.006437311414629221, + 0.005706117954105139, + 0.006858662236481905, + 0.007117889355868101, + 0.018504438921809196, + 0.008039021864533424, + 0.0142276706174016, + 0.01646740548312664, + 0.007618093863129616, + 0.029701607301831245, + 0.06436602771282196, + 0.01133294589817524, + 0.00555662764236331, + 0.004888809286057949, + 0.00579132791608572, + 0.00784284994006157, + 0.00846690684556961, + 0.008628347888588905, + 0.007266029249876738, + 0.010857176035642624, + 0.009656201116740704, + 0.016034498810768127, + 0.021912869065999985, + 0.016368700191378593, + 0.017416086047887802, + 0.01245360728353262, + 0.01752728410065174, + 0.017289169132709503, + 0.003925568889826536, + 0.008803321048617363, + 0.008346104063093662, + 0.00614843238145113, + 0.006371200084686279, + 0.005036453250795603, + 0.005424811039119959, + 0.003627133322879672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.00525687774643302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418695505708456 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.01418695505708456, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418695505708456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04511338099837303, + 0.03217301890254021, + 0.04248655587434769, + 0.07695233821868896, + 0.07990210503339767, + 0.04748574644327164, + 0.20489616692066193, + 0.15031148493289948, + 0.27864575386047363, + 0.042033471167087555 + ], + "credit_map": [ + 0.005633422639220953, + 0.003313241060823202, + 0.006690170615911484, + 0.006002491340041161, + 0.0046263583935797215, + 0.0066018300130963326, + 0.004329227842390537, + 0.007916638627648354, + 0.003416231367737055, + 0.00435212068259716, + 0.002548551419749856, + 0.004930880386382341, + 0.0046280017122626305, + 0.003113670041784644, + 0.004354661796241999, + 0.004828901030123234, + 0.005086459219455719, + 0.005366722121834755, + 0.003862903220579028, + 0.00420093908905983, + 0.007084147073328495, + 0.006273758132010698, + 0.004454470705240965, + 0.00615715654566884, + 0.009446453303098679, + 0.005089183803647757, + 0.008613471873104572, + 0.008990909904241562, + 0.01192127913236618, + 0.012086939997971058, + 0.006472683511674404, + 0.01433141902089119, + 0.00677111092954874, + 0.00753982225432992, + 0.012755095958709717, + 0.008615748025476933, + 0.009508858434855938, + 0.010990417562425137, + 0.006706222426146269, + 0.017014827579259872, + 0.004652251023799181, + 0.004817764274775982, + 0.004978839308023453, + 0.00495365634560585, + 0.00839599221944809, + 0.005836711265146732, + 0.00916807446628809, + 0.00468245567753911, + 0.020497912541031837, + 0.013594167307019234, + 0.02797364629805088, + 0.019731562584638596, + 0.016884226351976395, + 0.029631104320287704, + 0.057538460940122604, + 0.01904507912695408, + 0.017171302810311317, + 0.015107936225831509, + 0.012023396790027618, + 0.013983184471726418, + 0.015227554365992546, + 0.02732064202427864, + 0.01609869673848152, + 0.03337877616286278, + 0.02256825938820839, + 0.04420219734311104, + 0.054271142929792404, + 0.03380230814218521, + 0.040171898901462555, + 0.0279744490981102, + 0.03548813611268997, + 0.02016736939549446, + 0.0026992633938789368, + 0.005648657213896513, + 0.004926817025989294, + 0.0053368606604635715, + 0.00729580270126462, + 0.005066001787781715, + 0.007225162349641323, + 0.003834905568510294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.01418695505708456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011062951758503914 + ], + "opposite_logprob_tokens": [ + -4.511063098907471 + ], + "logprob_reward": -0.011062951758503914, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011062951758503914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04210209101438522, + 0.015189223922789097, + 0.03993208333849907, + 0.04234785959124565, + 0.04698722064495087, + 0.03678130730986595, + 0.3555179536342621, + 0.3197561800479889, + 0.0714702308177948, + 0.029915830120444298 + ], + "credit_map": [ + 0.005435023922473192, + 0.005127403885126114, + 0.008234016597270966, + 0.006104092579334974, + 0.0037869825027883053, + 0.004962037317454815, + 0.004671749658882618, + 0.0037807815242558718, + 0.0015115110436454415, + 0.0017056928481906652, + 0.0013825128553435206, + 0.0020168537739664316, + 0.0013823143672198057, + 0.002195715671405196, + 0.0017507430166006088, + 0.0032438801135867834, + 0.004306233488023281, + 0.007217286620289087, + 0.0038073209580034018, + 0.0041360859759151936, + 0.004053592216223478, + 0.006365480367094278, + 0.004540554713457823, + 0.005505530163645744, + 0.004323987755924463, + 0.004849803633987904, + 0.008439302444458008, + 0.0054249209351837635, + 0.004699346609413624, + 0.004897172097116709, + 0.00385738885961473, + 0.005855938885360956, + 0.0031696234364062548, + 0.0031556375324726105, + 0.0062878369353711605, + 0.0057950932532548904, + 0.0033294472377747297, + 0.007192826829850674, + 0.004896973725408316, + 0.01315978355705738, + 0.0034191457089036703, + 0.0029125267174094915, + 0.005079664755612612, + 0.004862753674387932, + 0.0039893873035907745, + 0.0052874828688800335, + 0.004128511529415846, + 0.007101832889020443, + 0.039449937641620636, + 0.019162792712450027, + 0.03167728707194328, + 0.03432304039597511, + 0.022361475974321365, + 0.05846596136689186, + 0.1239013597369194, + 0.026176081970334053, + 0.03762422502040863, + 0.023424390703439713, + 0.0260017067193985, + 0.09251797199249268, + 0.01652262918651104, + 0.029369227588176727, + 0.0735163614153862, + 0.020779641345143318, + 0.00497211841866374, + 0.00973824318498373, + 0.016010476276278496, + 0.010511698201298714, + 0.007839862257242203, + 0.00906302034854889, + 0.007311413995921612, + 0.0060234032571315765, + 0.002983124926686287, + 0.0037040752358734608, + 0.00424405001103878, + 0.004431282170116901, + 0.004702974576503038, + 0.0027142337057739496, + 0.004671517293900251, + 0.0024645705707371235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.011062951758503914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a Pegasus with a metallic blue body and cyan hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the Pegasus's mane and tail feathers display a vibrant rainbow spectrum with a neon inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the Pegasus perform a sequence of folding wings to dive diagonally downward and then unfurling them to ascend vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do long, brilliant trails composed of luminous particles emanate from the Pegasus's rainbow mane during its movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a low dynamic tracking angle and steadily rise to widen the field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862957075238228 + ], + "opposite_logprob_tokens": [ + -3.0486295223236084 + ], + "logprob_reward": -0.04862957075238228, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862957075238228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052181094884872437, + 0.025491613894701004, + 0.04540148749947548, + 0.07299958169460297, + 0.07142335176467896, + 0.03514306619763374, + 0.10812847316265106, + 0.48573464155197144, + 0.06778319180011749, + 0.03571348637342453 + ], + "credit_map": [ + 0.005832267459481955, + 0.005429319571703672, + 0.011914366856217384, + 0.005326816346496344, + 0.004882855340838432, + 0.007375419605523348, + 0.004681483842432499, + 0.006738574709743261, + 0.002588637638837099, + 0.003953429404646158, + 0.00261354586109519, + 0.0032407718244940042, + 0.002817578846588731, + 0.002861521439626813, + 0.002996396506205201, + 0.0044197323732078075, + 0.004611485172063112, + 0.008227754384279251, + 0.005760715808719397, + 0.00488959439098835, + 0.0054751234129071236, + 0.006617715582251549, + 0.003811570815742016, + 0.006007530726492405, + 0.00848209299147129, + 0.006137481890618801, + 0.009364132769405842, + 0.005908099003136158, + 0.01057872548699379, + 0.008342855609953403, + 0.0073118554428219795, + 0.016874339431524277, + 0.004575966391712427, + 0.006135361269116402, + 0.011791898868978024, + 0.007347316946834326, + 0.005584618542343378, + 0.010842793621122837, + 0.006289225537329912, + 0.018856173381209373, + 0.0026457214262336493, + 0.0023022517561912537, + 0.0044922311790287495, + 0.004068067763000727, + 0.0052245063707232475, + 0.003783550811931491, + 0.005861683748662472, + 0.006765058264136314, + 0.00912682618945837, + 0.007394477725028992, + 0.006612565368413925, + 0.009469345211982727, + 0.011588167399168015, + 0.018479302525520325, + 0.037442661821842194, + 0.008015133440494537, + 0.07173557579517365, + 0.03338095173239708, + 0.034372102469205856, + 0.11347302049398422, + 0.02812916971743107, + 0.046433817595243454, + 0.08607441931962967, + 0.07213561236858368, + 0.006266519892960787, + 0.008074595592916012, + 0.011570018716156483, + 0.009595728479325771, + 0.00918170902878046, + 0.008822519332170486, + 0.006711773108690977, + 0.007560333702713251, + 0.003214075695723295, + 0.006148653104901314, + 0.0052645765244960785, + 0.005272922571748495, + 0.004347861744463444, + 0.003717702580615878, + 0.00493120402097702, + 0.0028164908289909363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.04862957075238228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_tensor([4995, 4995])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_tensor([4995, 4995])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..71c79084a1dc09c8094b68f554e9647a3b6cf704 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_tensor([4995, 4995])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, high-angle, wide shot captures a sleek, long-range glider navigating a sky that appears not as a natural expanse but as a pure Abstract Expressionist canvas. The backdrop is a violent, emotional maelstrom of brushstrokes, where thick, cold cadmium red, deep cobalt blue, and acid yellow clash against a chaotic field of stark black and white. The glider, with its polished, pearlescent white fuselage and incredibly slender wings, traverses this turbulent, painted space in a slow, deliberate roll. The camera tracks steadily along the aircraft as it rotates around its central axis with calm, almost balletic precision. As the glider flips, its glossy surface captures and reflects chaotic streaks of pure color, its wings temporarily painting the sky with ever-shifting, aggressive patterns. Just as the roll reaches its peak, the entire perspective is shattered by a sudden, violent jolt; the camera strikes with a raw, physical force that momentarily blurs the frame, disrupts the fluid movement, introduces a visceral shock, and tears through the scene's fragile aerial tranquility.", + "seed": 11677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11677_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16634955694898962, + "vlm_logprob": -0.16634955694898962, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.16634955694898962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06374488025903702 + ], + "opposite_logprob_tokens": [ + -2.8137447834014893 + ], + "logprob_reward": -0.06374488025903702, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06374488025903702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.51441490650177, + 0.04128514230251312, + 0.029335161671042442, + 0.15246327221393585, + 0.021590199321508408, + 0.038714222609996796, + 0.12147045880556107, + 0.019538071006536484, + 0.01588107831776142, + 0.04530743882060051 + ], + "credit_map": [ + 0.10753294080495834, + 0.17971673607826233, + 0.02511548064649105, + 0.021088488399982452, + 0.11468430608510971, + 0.040358398109674454, + 0.01301037147641182, + 0.01290824729949236, + 0.0050990646705031395, + 0.003737658727914095, + 0.005161873996257782, + 0.005041389260441065, + 0.0055861580185592175, + 0.003996560350060463, + 0.008860033936798573, + 0.0038024086970835924, + 0.004054762423038483, + 0.003243508981540799, + 0.003416275605559349, + 0.002298877341672778, + 0.0028542266227304935, + 0.003815388074144721, + 0.004849995486438274, + 0.0048021296970546246, + 0.027219509705901146, + 0.047990571707487106, + 0.005010458640754223, + 0.008324000984430313, + 0.016785599291324615, + 0.03249713033437729, + 0.0076832533814013, + 0.006952761672437191, + 0.0028368926141411066, + 0.002339719794690609, + 0.004853225778788328, + 0.0018715413752943277, + 0.0014906016876921058, + 0.0019439266761764884, + 0.004255765583366156, + 0.001998527441173792, + 0.003109056269749999, + 0.0032133881468325853, + 0.004039364401251078, + 0.0062374211847782135, + 0.004423238802701235, + 0.0032833809964358807, + 0.009177724830806255, + 0.005230647511780262, + 0.02679968811571598, + 0.013432743027806282, + 0.013132292777299881, + 0.010737373493611813, + 0.01694752834737301, + 0.013834089040756226, + 0.009965711273252964, + 0.016621045768260956, + 0.0018951999954879284, + 0.0012779447715729475, + 0.0019972214940935373, + 0.0027869869954884052, + 0.002528380835428834, + 0.0027384182903915644, + 0.003082703333348036, + 0.003231218084692955, + 0.0017135848756879568, + 0.0015067849308252335, + 0.002628972986713052, + 0.002055496210232377, + 0.0019067421089857817, + 0.001614737673662603, + 0.001879477989859879, + 0.0025752822402864695, + 0.004833908285945654, + 0.008529824204742908, + 0.0043021864257752895, + 0.007980256341397762, + 0.004949996247887611, + 0.006133438553661108, + 0.00401155324652791, + 0.004566278774291277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.06374488025903702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262214224785566 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.005262214224785566, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262214224785566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06361691653728485, + 0.07609762251377106, + 0.04444292187690735, + 0.2064874917268753, + 0.022379614412784576, + 0.033640045672655106, + 0.1018919125199318, + 0.0160970538854599, + 0.010119592770934105, + 0.42522692680358887 + ], + "credit_map": [ + 0.0167998019605875, + 0.012100578285753727, + 0.005567561835050583, + 0.004871631506830454, + 0.00978995580226183, + 0.007145296316593885, + 0.0038316536229103804, + 0.0035104311536997557, + 0.009299339726567268, + 0.009859560057520866, + 0.0076070926152169704, + 0.014759176410734653, + 0.006836744956672192, + 0.007632069289684296, + 0.01243453100323677, + 0.0076690963469445705, + 0.005852935835719109, + 0.005032150074839592, + 0.005407453514635563, + 0.003922128118574619, + 0.004083481151610613, + 0.0056479135528206825, + 0.006486271973699331, + 0.008010582998394966, + 0.037562351673841476, + 0.06654559820890427, + 0.006349501665681601, + 0.011239815503358841, + 0.02297893911600113, + 0.04342574626207352, + 0.008859449066221714, + 0.009526071138679981, + 0.002612828044220805, + 0.003167687216773629, + 0.006700910162180662, + 0.0018579716561362147, + 0.0019339704886078835, + 0.0016448319656774402, + 0.0020356085151433945, + 0.002425804967060685, + 0.0027909097261726856, + 0.0026911641471087933, + 0.003313134191557765, + 0.004502609837800264, + 0.005460216198116541, + 0.0028946958482265472, + 0.007186896167695522, + 0.004800416063517332, + 0.021000787615776062, + 0.008326290175318718, + 0.010212493129074574, + 0.014270246960222721, + 0.014099907130002975, + 0.011829675175249577, + 0.009547103196382523, + 0.012605398893356323, + 0.0018086766358464956, + 0.0013420555042102933, + 0.0014124623266980052, + 0.0030734450556337833, + 0.0020719754975289106, + 0.0027157512959092855, + 0.0017496966756880283, + 0.0019229904282838106, + 0.0011132138315588236, + 0.0008664608467370272, + 0.0014323045033961535, + 0.0015844933222979307, + 0.0013044556835666299, + 0.0014126638416200876, + 0.0014789820415899158, + 0.0009270183509215713, + 0.021631790325045586, + 0.008124695159494877, + 0.08905066549777985, + 0.043529096990823746, + 0.011966381222009659, + 0.03613795340061188, + 0.05420155078172684, + 0.16058474779129028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.005262214224785566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987939715385437 + ], + "opposite_logprob_tokens": [ + -1.8048794269561768 + ], + "logprob_reward": -0.17987939715385437, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987939715385437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23503762483596802, + 0.038611263036727905, + 0.024373432621359825, + 0.15616875886917114, + 0.05144855007529259, + 0.03597157448530197, + 0.11745645850896835, + 0.021698318421840668, + 0.035511597990989685, + 0.2837224304676056 + ], + "credit_map": [ + 0.08301087468862534, + 0.06878285855054855, + 0.010460089892148972, + 0.01214046124368906, + 0.021930571645498276, + 0.026365084573626518, + 0.006110675632953644, + 0.006237013265490532, + 0.005196784622967243, + 0.0032378495670855045, + 0.004691444803029299, + 0.0048566958867013454, + 0.003765429137274623, + 0.004231908358633518, + 0.006394665688276291, + 0.00623648427426815, + 0.004897115286439657, + 0.0030228784307837486, + 0.0020423911046236753, + 0.0019101510988548398, + 0.0021002772264182568, + 0.002882142784073949, + 0.00366711663082242, + 0.0038513620384037495, + 0.024577470496296883, + 0.041043397039175034, + 0.0068422467447817326, + 0.00914586428552866, + 0.02042367309331894, + 0.03445480763912201, + 0.011052947491407394, + 0.008628344163298607, + 0.005406321492046118, + 0.008535097353160381, + 0.006835966370999813, + 0.005712429061532021, + 0.0052147493697702885, + 0.004313085228204727, + 0.00787106342613697, + 0.0075598363764584064, + 0.003177753183990717, + 0.0034939711913466454, + 0.00468665175139904, + 0.004495145287364721, + 0.006906963884830475, + 0.00379230547696352, + 0.005112317390739918, + 0.00430646538734436, + 0.029591763392090797, + 0.011419498361647129, + 0.011679885908961296, + 0.017144517973065376, + 0.012677549384534359, + 0.011688929051160812, + 0.008893724530935287, + 0.0143605787307024, + 0.00307927536778152, + 0.0017387415282428265, + 0.002083358820527792, + 0.0037468355149030685, + 0.0019256708910688758, + 0.0029123842250555754, + 0.002280634129419923, + 0.003931417595595121, + 0.0033538760617375374, + 0.003956654109060764, + 0.005864162463694811, + 0.005625319201499224, + 0.0038857392501085997, + 0.00362241268157959, + 0.004185807891190052, + 0.005017626564949751, + 0.013748792931437492, + 0.007456441875547171, + 0.05920308455824852, + 0.028490057215094566, + 0.008675667457282543, + 0.023766860365867615, + 0.03620144724845886, + 0.10618007928133011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.17987939715385437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14002880454063416, + 0.07183001190423965, + 0.02262282744050026, + 0.18501931428909302, + 0.023239530622959137, + 0.026571327820420265, + 0.053723473101854324, + 0.022242799401283264, + 0.011299480684101582, + 0.44342246651649475 + ], + "credit_map": [ + 0.04527071863412857, + 0.019766628742218018, + 0.009766928851604462, + 0.009209255687892437, + 0.029762281104922295, + 0.013923865742981434, + 0.006594667211174965, + 0.005734457168728113, + 0.012742363847792149, + 0.009074583649635315, + 0.005522013641893864, + 0.014781538397073746, + 0.007322820834815502, + 0.0059602390974760056, + 0.01041175052523613, + 0.0060146949253976345, + 0.0037089530378580093, + 0.002876016777008772, + 0.002288923831656575, + 0.001983320340514183, + 0.002613931195810437, + 0.002851194003596902, + 0.0034767119213938713, + 0.0028237730730324984, + 0.03333079442381859, + 0.05857507511973381, + 0.005994338076561689, + 0.009430891834199429, + 0.020646095275878906, + 0.039490263909101486, + 0.009136218577623367, + 0.008415617980062962, + 0.0027022624853998423, + 0.003516207914799452, + 0.004179134499281645, + 0.0025786173064261675, + 0.0024957400746643543, + 0.0019354976247996092, + 0.0031328776385635138, + 0.0026991916820406914, + 0.002177122514694929, + 0.0028612217865884304, + 0.002608230570331216, + 0.003208656795322895, + 0.0032468363642692566, + 0.00201974855735898, + 0.007077844347804785, + 0.0033716661855578423, + 0.011280803941190243, + 0.005347502883523703, + 0.004765476565808058, + 0.00841530691832304, + 0.005929193459451199, + 0.006095541641116142, + 0.004757003393024206, + 0.007132640574127436, + 0.004169638734310865, + 0.0022997099440544844, + 0.002416380448266864, + 0.0030411886982619762, + 0.0018058615969493985, + 0.0024609833490103483, + 0.0025100279599428177, + 0.003539007157087326, + 0.0014970812480896711, + 0.0012750320602208376, + 0.0018906703917309642, + 0.0012679527280852199, + 0.0014730937546119094, + 0.0012101596221327782, + 0.0011238495353609324, + 0.0015616407617926598, + 0.023023653775453568, + 0.010804034769535065, + 0.09141038358211517, + 0.04682525619864464, + 0.012509784661233425, + 0.03661983832716942, + 0.05633372440934181, + 0.1658957451581955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0067422278225421906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576119065284729 + ], + "opposite_logprob_tokens": [ + -0.826119065284729 + ], + "logprob_reward": -0.576119065284729, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576119065284729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17846210300922394, + 0.06388575583696365, + 0.05058974027633667, + 0.29307830333709717, + 0.05508111044764519, + 0.07833565771579742, + 0.08551477640867233, + 0.03071998618543148, + 0.060862548649311066, + 0.10346999764442444 + ], + "credit_map": [ + 0.05554706230759621, + 0.0415782630443573, + 0.01115841418504715, + 0.012927255593240261, + 0.02076987735927105, + 0.018002165481448174, + 0.00820782408118248, + 0.010271251201629639, + 0.0088365962728858, + 0.009749554097652435, + 0.006786154583096504, + 0.007805100176483393, + 0.006438056938350201, + 0.0066590690985322, + 0.010511795058846474, + 0.007099431473761797, + 0.007469581440091133, + 0.006616387516260147, + 0.005807938054203987, + 0.0040580108761787415, + 0.005360769107937813, + 0.006122291088104248, + 0.006795166525989771, + 0.008359599858522415, + 0.05232212319970131, + 0.09686978906393051, + 0.01690402440726757, + 0.01818888820707798, + 0.024101804941892624, + 0.04146882891654968, + 0.027859151363372803, + 0.015363700687885284, + 0.008220487274229527, + 0.011534933000802994, + 0.008116955868899822, + 0.0054529220797121525, + 0.0062331692315638065, + 0.005691941361874342, + 0.00466366158798337, + 0.005167035851627588, + 0.005906166974455118, + 0.006568946409970522, + 0.007863416336476803, + 0.011233680881559849, + 0.0104269003495574, + 0.007237469311803579, + 0.01922193355858326, + 0.009877143427729607, + 0.019723372533917427, + 0.010156597010791302, + 0.006680808495730162, + 0.022567493841052055, + 0.008052844554185867, + 0.0056198700331151485, + 0.006344831082969904, + 0.006368961185216904, + 0.005700086709111929, + 0.0029890937730669975, + 0.003765235422179103, + 0.004440443590283394, + 0.0024893758818507195, + 0.003705965355038643, + 0.0035755711141973734, + 0.004054215271025896, + 0.005177129991352558, + 0.006361895706504583, + 0.009298989549279213, + 0.009055373258888721, + 0.008704918436706066, + 0.006999582052230835, + 0.006779839750379324, + 0.00848482083529234, + 0.009653269313275814, + 0.011065501719713211, + 0.015043201856315136, + 0.009025930427014828, + 0.012451743707060814, + 0.016072258353233337, + 0.010789952240884304, + 0.019368138164281845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.576119065284729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle, wide shot captures a sleek, long-range glider navigating a sky that appears not as a natural expanse but as a pure Abstract Expressionist canvas. The backdrop is a violent, emotional maelstrom of brushstrokes, where thick, cold cadmium red, deep cobalt blue, and acid yellow clash against a chaotic field of stark black and white. The glider, with its polished, pearlescent white fuselage and incredibly slender wings, traverses this turbulent, painted space in a slow, deliberate roll. The camera tracks steadily along the aircraft as it rotates around its central axis with calm, almost balletic precision. As the glider flips, its glossy surface captures and reflects chaotic streaks of pure color, its wings temporarily painting the sky with ever-shifting, aggressive patterns. Just as the roll reaches its peak, the entire perspective is shattered by a sudden, violent jolt; the camera strikes with a raw, physical force that momentarily blurs the frame, disrupts the fluid movement, introduces a visceral shock, and tears through the scene's fragile aerial tranquility.", + "seed": 11678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11678_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12953550536185504, + "vlm_logprob": -0.12953550536185504, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.12953550536185504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28285348415374756 + ], + "opposite_logprob_tokens": [ + -1.4078534841537476 + ], + "logprob_reward": -0.28285348415374756, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28285348415374756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08310646563768387, + 0.13600923120975494, + 0.0747683197259903, + 0.1462489515542984, + 0.09988328814506531, + 0.15171995759010315, + 0.09925973415374756, + 0.043178509920835495, + 0.09914660453796387, + 0.066678985953331 + ], + "credit_map": [ + 0.010039114393293858, + 0.014733601361513138, + 0.010828744620084763, + 0.01018916442990303, + 0.009979154914617538, + 0.008400430902838707, + 0.009784383699297905, + 0.009151875972747803, + 0.01590585522353649, + 0.011137725785374641, + 0.048720020800828934, + 0.01844171993434429, + 0.00797904934734106, + 0.010096823796629906, + 0.008545287884771824, + 0.015182745642960072, + 0.008733788505196571, + 0.009927203878760338, + 0.012410431168973446, + 0.009697061963379383, + 0.009116437286138535, + 0.009559166617691517, + 0.007018909323960543, + 0.0083053233101964, + 0.023393599316477776, + 0.022908516228199005, + 0.014282974414527416, + 0.016963854432106018, + 0.01908099465072155, + 0.012089400552213192, + 0.01777399145066738, + 0.01975562609732151, + 0.015211226418614388, + 0.01557803712785244, + 0.008758236654102802, + 0.010562662035226822, + 0.019000748172402382, + 0.011842737905681133, + 0.009126204065978527, + 0.009803436696529388, + 0.02382699027657509, + 0.023405106738209724, + 0.015013408847153187, + 0.018287835642695427, + 0.019873449578881264, + 0.012432605028152466, + 0.020424859598279, + 0.018455693498253822, + 0.007834631018340588, + 0.009548511356115341, + 0.00972316600382328, + 0.009668787941336632, + 0.01487218402326107, + 0.017161255702376366, + 0.02299705520272255, + 0.007454147562384605, + 0.003907081205397844, + 0.004735409747809172, + 0.004181951750069857, + 0.004648514091968536, + 0.004904576111584902, + 0.006815242115408182, + 0.007974531501531601, + 0.006011202000081539, + 0.010030888020992279, + 0.011932938359677792, + 0.021838350221514702, + 0.012943324632942677, + 0.011409994214773178, + 0.011337393894791603, + 0.010494724847376347, + 0.009158985689282417, + 0.006146787200123072, + 0.01033980306237936, + 0.010013213381171227, + 0.009433615952730179, + 0.007333662826567888, + 0.007223520427942276, + 0.00885129813104868, + 0.007337082643061876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.28285348415374756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748621817678213 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748621817678213, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748621817678213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1329236924648285, + 0.08669691532850266, + 0.05263340845704079, + 0.3097631335258484, + 0.08042124658823013, + 0.05087946727871895, + 0.026449671015143394, + 0.17802749574184418, + 0.039266422390937805, + 0.04293861612677574 + ], + "credit_map": [ + 0.01935192011296749, + 0.02010360173881054, + 0.010764248669147491, + 0.0146050825715065, + 0.019653722643852234, + 0.015205392614006996, + 0.0118552939966321, + 0.02138441801071167, + 0.011678967624902725, + 0.009904770180583, + 0.01750711165368557, + 0.007548886816948652, + 0.0090060168877244, + 0.015210089273750782, + 0.007490473799407482, + 0.008350602351129055, + 0.006299797911196947, + 0.00627814419567585, + 0.007804040797054768, + 0.008289216086268425, + 0.006631271447986364, + 0.007446340285241604, + 0.004445130471140146, + 0.005439466331154108, + 0.021879585459828377, + 0.05079672485589981, + 0.01955290324985981, + 0.028143687173724174, + 0.028004435822367668, + 0.028163615614175797, + 0.03929428383708, + 0.09392788261175156, + 0.012293785810470581, + 0.011973680928349495, + 0.008275733329355717, + 0.009628640487790108, + 0.014236376620829105, + 0.00914044864475727, + 0.007694633677601814, + 0.007177941035479307, + 0.009200806729495525, + 0.007702615577727556, + 0.0048837303183972836, + 0.00533012580126524, + 0.007860036566853523, + 0.006370462477207184, + 0.004951683804392815, + 0.0045800055377185345, + 0.002646335167810321, + 0.0031634815968573093, + 0.0024911987129598856, + 0.0020569805055856705, + 0.0038838537875562906, + 0.0054017649963498116, + 0.00407791743054986, + 0.002728138817474246, + 0.027790017426013947, + 0.020063288509845734, + 0.028272708877921104, + 0.017614757642149925, + 0.01915988326072693, + 0.016958652064204216, + 0.027504311874508858, + 0.020663874223828316, + 0.0028301626443862915, + 0.0032118118833750486, + 0.005333660636097193, + 0.005135256797075272, + 0.005654593463987112, + 0.004598916508257389, + 0.007905501872301102, + 0.004596515092998743, + 0.0036194445565342903, + 0.007712372113019228, + 0.008791608735918999, + 0.004793636500835419, + 0.004600249696522951, + 0.003632602049037814, + 0.004555465187877417, + 0.005233237519860268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.006748621817678213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813188433647156 + ], + "opposite_logprob_tokens": [ + -1.4063189029693604 + ], + "logprob_reward": -0.2813188433647156, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813188433647156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07739309966564178, + 0.20757117867469788, + 0.1355205625295639, + 0.09849876910448074, + 0.10316578298807144, + 0.07594919204711914, + 0.038550589233636856, + 0.08498450368642807, + 0.08928143978118896, + 0.08908485621213913 + ], + "credit_map": [ + 0.008737959899008274, + 0.006964104250073433, + 0.010253369808197021, + 0.010043005459010601, + 0.009325086139142513, + 0.007903206162154675, + 0.015821155160665512, + 0.00834521371871233, + 0.025393236428499222, + 0.014180583879351616, + 0.08589070290327072, + 0.025834232568740845, + 0.010249282233417034, + 0.012703406624495983, + 0.012364253401756287, + 0.02095545455813408, + 0.02150730788707733, + 0.020085670053958893, + 0.018349656835198402, + 0.02795681729912758, + 0.010002785362303257, + 0.01420445367693901, + 0.010947548784315586, + 0.012466329149901867, + 0.011148673482239246, + 0.01420536357909441, + 0.013116752728819847, + 0.011939942836761475, + 0.009117168374359608, + 0.010966205969452858, + 0.00882662832736969, + 0.019178038462996483, + 0.01399625651538372, + 0.012842297554016113, + 0.011266862973570824, + 0.017978042364120483, + 0.012636451050639153, + 0.009717580862343311, + 0.01410932932049036, + 0.010618962347507477, + 0.01297521311789751, + 0.011672778986394405, + 0.008004601113498211, + 0.007495776284486055, + 0.01031047198921442, + 0.00673131225630641, + 0.008856882341206074, + 0.009902152232825756, + 0.004276918713003397, + 0.0035489366855472326, + 0.005236795172095299, + 0.004339015576988459, + 0.005179841071367264, + 0.005562290549278259, + 0.0071616582572460175, + 0.0032451350707560778, + 0.010275458917021751, + 0.008249563165009022, + 0.009187804535031319, + 0.009287442080676556, + 0.009308114647865295, + 0.008740867488086224, + 0.021169211715459824, + 0.008766047656536102, + 0.007605310995131731, + 0.008279082365334034, + 0.014936055988073349, + 0.010526763275265694, + 0.0131442965939641, + 0.010391966439783573, + 0.015125475823879242, + 0.00927248690277338, + 0.01007295586168766, + 0.009582646191120148, + 0.018538063392043114, + 0.00982914213091135, + 0.007879996672272682, + 0.008845540694892406, + 0.008349304087460041, + 0.015987204387784004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.2813188433647156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760250985622406, + 0.09647457301616669, + 0.058747876435518265, + 0.2508307993412018, + 0.07563511282205582, + 0.10201402753591537, + 0.0688970610499382, + 0.028574664145708084, + 0.13939744234085083, + 0.10340336710214615 + ], + "credit_map": [ + 0.0077575622126460075, + 0.00882689468562603, + 0.008554534986615181, + 0.012718192301690578, + 0.009532266296446323, + 0.006060989107936621, + 0.011747407726943493, + 0.010827247053384781, + 0.011071493849158287, + 0.009598291479051113, + 0.02768651396036148, + 0.01296543050557375, + 0.007614079397171736, + 0.008108013309538364, + 0.007648843340575695, + 0.011781904846429825, + 0.00797534454613924, + 0.00838532391935587, + 0.007590115070343018, + 0.008548881858587265, + 0.007509704679250717, + 0.006740543991327286, + 0.00574678648263216, + 0.0062511758878827095, + 0.0350823774933815, + 0.04139787331223488, + 0.02005881257355213, + 0.02515697106719017, + 0.026007873937487602, + 0.01776539720594883, + 0.026209937408566475, + 0.0591515451669693, + 0.012928026728332043, + 0.011675305664539337, + 0.007006642408668995, + 0.010631703771650791, + 0.011902754195034504, + 0.008137076161801815, + 0.00681516295298934, + 0.006538437679409981, + 0.015774279832839966, + 0.01539729442447424, + 0.010422100313007832, + 0.011072585359215736, + 0.013591621071100235, + 0.008968711830675602, + 0.013885919004678726, + 0.012901516631245613, + 0.005543015897274017, + 0.007658810820430517, + 0.006471320521086454, + 0.007407369092106819, + 0.008903958834707737, + 0.009145503863692284, + 0.018302440643310547, + 0.005464639514684677, + 0.003240027232095599, + 0.004830357618629932, + 0.002802947536110878, + 0.002582504181191325, + 0.003569895401597023, + 0.004062884021550417, + 0.0034938794560730457, + 0.003992168698459864, + 0.009215139783918858, + 0.01132748369127512, + 0.02150029130280018, + 0.016217079013586044, + 0.021152062341570854, + 0.017252236604690552, + 0.0282549187541008, + 0.014478224329650402, + 0.01014458667486906, + 0.017317309975624084, + 0.018778717145323753, + 0.013221532106399536, + 0.010703272186219692, + 0.010569963604211807, + 0.013302739709615707, + 0.009365245699882507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.006740096490830183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001648098230362 + ], + "opposite_logprob_tokens": [ + -2.695016384124756 + ], + "logprob_reward": -0.07001648098230362, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001648098230362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06611497700214386, + 0.09353067725896835, + 0.18058262765407562, + 0.06262698769569397, + 0.05605703964829445, + 0.05623244121670723, + 0.0636916384100914, + 0.11613285541534424, + 0.13653038442134857, + 0.1685003638267517 + ], + "credit_map": [ + 0.008036619983613491, + 0.006231013685464859, + 0.007622492033988237, + 0.010227902792394161, + 0.00900264922529459, + 0.005949328653514385, + 0.008669223636388779, + 0.010375747457146645, + 0.011267454363405704, + 0.009647385217249393, + 0.02110281214118004, + 0.011417257599532604, + 0.008620609529316425, + 0.01018893625587225, + 0.009358304552733898, + 0.011927912011742592, + 0.02733183652162552, + 0.025747500360012054, + 0.03308722376823425, + 0.04197019711136818, + 0.013318097218871117, + 0.015388001687824726, + 0.011285427957773209, + 0.012454350478947163, + 0.00765680568292737, + 0.009308927692472935, + 0.007726422976702452, + 0.008205253630876541, + 0.00543915294110775, + 0.006484265439212322, + 0.0067018987610936165, + 0.011104266159236431, + 0.008709508925676346, + 0.008360509760677814, + 0.0076735783368349075, + 0.00686988839879632, + 0.0078022838570177555, + 0.006017755251377821, + 0.0054179285652935505, + 0.005205577705055475, + 0.007260662969201803, + 0.007574766408652067, + 0.008009343408048153, + 0.007824186235666275, + 0.00604264298453927, + 0.0061779017560184, + 0.0067045604810118675, + 0.006638378370553255, + 0.006415033712983131, + 0.008923596702516079, + 0.006349433213472366, + 0.006189580075442791, + 0.00881279818713665, + 0.008688990026712418, + 0.011488445103168488, + 0.006823759060353041, + 0.008788343518972397, + 0.015902802348136902, + 0.008415787480771542, + 0.01150426734238863, + 0.014307905919849873, + 0.015068666078150272, + 0.030002500861883163, + 0.012142583727836609, + 0.007055709138512611, + 0.009296149015426636, + 0.027141254395246506, + 0.023506920784711838, + 0.012957293540239334, + 0.0181844774633646, + 0.025752389803528786, + 0.012636189348995686, + 0.018047628924250603, + 0.02679738588631153, + 0.031000318005681038, + 0.02109992504119873, + 0.01624661684036255, + 0.021330831572413445, + 0.018883857876062393, + 0.015093800611793995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.07001648098230362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle, wide shot captures a sleek, long-range glider navigating a sky that appears not as a natural expanse but as a pure Abstract Expressionist canvas. The backdrop is a violent, emotional maelstrom of brushstrokes, where thick, cold cadmium red, deep cobalt blue, and acid yellow clash against a chaotic field of stark black and white. The glider, with its polished, pearlescent white fuselage and incredibly slender wings, traverses this turbulent, painted space in a slow, deliberate roll. The camera tracks steadily along the aircraft as it rotates around its central axis with calm, almost balletic precision. As the glider flips, its glossy surface captures and reflects chaotic streaks of pure color, its wings temporarily painting the sky with ever-shifting, aggressive patterns. Just as the roll reaches its peak, the entire perspective is shattered by a sudden, violent jolt; the camera strikes with a raw, physical force that momentarily blurs the frame, disrupts the fluid movement, introduces a visceral shock, and tears through the scene's fragile aerial tranquility.", + "seed": 11679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11679_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10173438466154039, + "vlm_logprob": -0.10173438466154039, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.10173438466154039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025479525327682495 + ], + "opposite_logprob_tokens": [ + -3.775479555130005 + ], + "logprob_reward": -0.025479525327682495, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025479525327682495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07190801948308945, + 0.029936542734503746, + 0.03336525708436966, + 0.12955287098884583, + 0.03356017544865608, + 0.09581930935382843, + 0.07979287952184677, + 0.09739954769611359, + 0.2434229701757431, + 0.1852424591779709 + ], + "credit_map": [ + 0.0090039586648345, + 0.006284231785684824, + 0.009386438876390457, + 0.012235036119818687, + 0.009430664591491222, + 0.008230234496295452, + 0.007312202826142311, + 0.010025253519415855, + 0.007671893108636141, + 0.0039014769718050957, + 0.0031796612311154604, + 0.0017329438123852015, + 0.0028047668747603893, + 0.0061688413843512535, + 0.001664704061113298, + 0.002812255872413516, + 0.005710418801754713, + 0.008163322694599628, + 0.004042471759021282, + 0.0032399632036685944, + 0.0034117880277335644, + 0.0035367191303521395, + 0.00226088659837842, + 0.0029996875673532486, + 0.030941065400838852, + 0.008343115448951721, + 0.007776192389428616, + 0.004908118862658739, + 0.04522404074668884, + 0.016335388645529747, + 0.007040602620691061, + 0.008984342217445374, + 0.004594962578266859, + 0.005258095450699329, + 0.0040170359425246716, + 0.003906975965946913, + 0.005083984229713678, + 0.0034551043063402176, + 0.002834948943927884, + 0.004409069661051035, + 0.01164774876087904, + 0.014836148358881474, + 0.01072053611278534, + 0.0080553675070405, + 0.012953415513038635, + 0.013568485155701637, + 0.009393680840730667, + 0.014643928036093712, + 0.017447566613554955, + 0.015197192318737507, + 0.007358622271567583, + 0.008871142752468586, + 0.008590711280703545, + 0.00841995608061552, + 0.007762168068438768, + 0.006145521067082882, + 0.012304764240980148, + 0.013707440346479416, + 0.012448733672499657, + 0.009922325611114502, + 0.01239390391856432, + 0.015854276716709137, + 0.010944034904241562, + 0.009824061766266823, + 0.027928069233894348, + 0.046797141432762146, + 0.03443930298089981, + 0.039153702557086945, + 0.030906856060028076, + 0.023139946162700653, + 0.018743325024843216, + 0.022314608097076416, + 0.03652390465140343, + 0.020743682980537415, + 0.038502514362335205, + 0.028639759868383408, + 0.01581854559481144, + 0.013810261152684689, + 0.018536891788244247, + 0.012666894122958183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.025479525327682495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764184445142746 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.00764184445142746, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764184445142746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0941319614648819, + 0.07785897701978683, + 0.07416407018899918, + 0.16989418864250183, + 0.09158042073249817, + 0.2016308754682541, + 0.09443933516740799, + 0.045778702944517136, + 0.0797896534204483, + 0.0707317516207695 + ], + "credit_map": [ + 0.011611766181886196, + 0.007444905117154121, + 0.010136628523468971, + 0.01523287408053875, + 0.01457243598997593, + 0.010119476355612278, + 0.012952746823430061, + 0.01206114236265421, + 0.016683204099535942, + 0.011665711179375648, + 0.010746748186647892, + 0.005082009360194206, + 0.008135026320815086, + 0.010789306834340096, + 0.005484193563461304, + 0.009272786788642406, + 0.013920366764068604, + 0.013616894371807575, + 0.015272997319698334, + 0.006020115688443184, + 0.007018157280981541, + 0.006948911119252443, + 0.004979530815035105, + 0.006387113127857447, + 0.043059591203927994, + 0.011652505025267601, + 0.008075841702520847, + 0.004277431406080723, + 0.06218394264578819, + 0.021536799147725105, + 0.007129068952053785, + 0.011979038827121258, + 0.012426331639289856, + 0.012582858093082905, + 0.012309050187468529, + 0.010323639959096909, + 0.010202346369624138, + 0.011084873229265213, + 0.008553173393011093, + 0.014098163694143295, + 0.02367607317864895, + 0.031150875613093376, + 0.01607460156083107, + 0.012182886712253094, + 0.020961156114935875, + 0.032511256635189056, + 0.012215975672006607, + 0.052858076989650726, + 0.018060455098748207, + 0.015936383977532387, + 0.010511383414268494, + 0.010171115398406982, + 0.011498695239424706, + 0.01371646486222744, + 0.00712454691529274, + 0.007420305162668228, + 0.00761193223297596, + 0.00887877494096756, + 0.006160310469567776, + 0.005920371040701866, + 0.0046751732006669044, + 0.004081223625689745, + 0.003989654127508402, + 0.00446127075701952, + 0.010094234719872475, + 0.0173994991928339, + 0.009692016057670116, + 0.016491344198584557, + 0.007489432115107775, + 0.005654120817780495, + 0.006502995733171701, + 0.00646601989865303, + 0.011289906688034534, + 0.007883713580667973, + 0.009271047078073025, + 0.012527693063020706, + 0.008524800650775433, + 0.006851930171251297, + 0.009128434583544731, + 0.005254238843917847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.00764184445142746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305974766612053 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305974766612053, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305974766612053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08229045569896698, + 0.046668075025081635, + 0.020344499498605728, + 0.1184186115860939, + 0.033172834664583206, + 0.05504906550049782, + 0.10808519273996353, + 0.10457583516836166, + 0.10324449837207794, + 0.3281509280204773 + ], + "credit_map": [ + 0.009793451055884361, + 0.007907157763838768, + 0.01062812004238367, + 0.01458149217069149, + 0.010600311681628227, + 0.008614836260676384, + 0.008730598725378513, + 0.011434483341872692, + 0.007215197663754225, + 0.007235257420688868, + 0.004207867197692394, + 0.00460815941914916, + 0.006859495770186186, + 0.005590200889855623, + 0.004535938613116741, + 0.006415956653654575, + 0.003980844747275114, + 0.0026255843695253134, + 0.0023193196393549442, + 0.002036820398643613, + 0.0025954449083656073, + 0.0028087825048714876, + 0.0015332431066781282, + 0.002444459591060877, + 0.026126552373170853, + 0.008699887432157993, + 0.006223473697900772, + 0.005146828014403582, + 0.044330939650535583, + 0.013851267285645008, + 0.005036345683038235, + 0.009003317914903164, + 0.004219899419695139, + 0.005003608763217926, + 0.0028553444426506758, + 0.0028704172000288963, + 0.007616141345351934, + 0.00324276858009398, + 0.0027358883526176214, + 0.004628767259418964, + 0.008751877583563328, + 0.008641046471893787, + 0.006346135400235653, + 0.005303774494677782, + 0.005090735852718353, + 0.008450549095869064, + 0.004801648203283548, + 0.007663297932595015, + 0.015839247032999992, + 0.01815142296254635, + 0.009431419894099236, + 0.009722764603793621, + 0.012142769061028957, + 0.020495934411883354, + 0.009940455667674541, + 0.012361167930066586, + 0.01678900606930256, + 0.018091384321451187, + 0.017343800514936447, + 0.012756426818668842, + 0.00957455299794674, + 0.008978407829999924, + 0.011470700614154339, + 0.009571553207933903, + 0.013495370745658875, + 0.022004544734954834, + 0.014101113192737103, + 0.023664232343435287, + 0.009078538976609707, + 0.0068502891808748245, + 0.008326200768351555, + 0.0057242149487137794, + 0.026552071794867516, + 0.026464974507689476, + 0.060551319271326065, + 0.1265294998884201, + 0.015339686535298824, + 0.013431979343295097, + 0.035822510719299316, + 0.023458871990442276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.04305974766612053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06721227616071701, + 0.054696328938007355, + 0.06159135699272156, + 0.263249933719635, + 0.04423842579126358, + 0.05157007649540901, + 0.07718978822231293, + 0.09472723305225372, + 0.1295911967754364, + 0.15593339502811432 + ], + "credit_map": [ + 0.006289007142186165, + 0.007186579052358866, + 0.008165895938873291, + 0.011619840748608112, + 0.01065752562135458, + 0.006004499737173319, + 0.007555090822279453, + 0.009733838029205799, + 0.010427691042423248, + 0.007922948338091373, + 0.006375482305884361, + 0.0034907436929643154, + 0.005441790912300348, + 0.012337584048509598, + 0.003950096666812897, + 0.00474999425932765, + 0.009254366159439087, + 0.007543155457824469, + 0.018474701792001724, + 0.00481544341892004, + 0.003893477376550436, + 0.01035990472882986, + 0.003981131594628096, + 0.0032691792584955692, + 0.05886039510369301, + 0.016150547191500664, + 0.01715114898979664, + 0.010422894731163979, + 0.09713917970657349, + 0.031455397605895996, + 0.013839141465723515, + 0.018231241032481194, + 0.005729747470468283, + 0.006461296230554581, + 0.005268233362585306, + 0.005287846550345421, + 0.008634567260742188, + 0.004791997838765383, + 0.0035067982971668243, + 0.004557937383651733, + 0.00841135811060667, + 0.008391639217734337, + 0.0053229681216180325, + 0.003966281656175852, + 0.008135216310620308, + 0.007573009934276342, + 0.004130913410335779, + 0.005638688337057829, + 0.010068932548165321, + 0.012050669640302658, + 0.00585605762898922, + 0.006712593603879213, + 0.0102836973965168, + 0.018399709835648537, + 0.0056119137443602085, + 0.008206218481063843, + 0.013691521249711514, + 0.014082955196499825, + 0.012392775155603886, + 0.012248030863702297, + 0.009352579712867737, + 0.00862930528819561, + 0.010499437339603901, + 0.013830626383423805, + 0.01182472426444292, + 0.01506955549120903, + 0.012868187390267849, + 0.02069968171417713, + 0.017321737483143806, + 0.01966143772006035, + 0.015529721975326538, + 0.01661614514887333, + 0.0146113820374012, + 0.0129928607493639, + 0.02210107073187828, + 0.05291453376412392, + 0.01216504629701376, + 0.009036489762365818, + 0.01681441254913807, + 0.015297616831958294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.0036162238102406263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42887458205223083 + ], + "opposite_logprob_tokens": [ + -1.0538746118545532 + ], + "logprob_reward": -0.42887458205223083, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42887458205223083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023775024339556694, + 0.016276953741908073, + 0.023543380200862885, + 0.037959132343530655, + 0.03873955085873604, + 0.053766388446092606, + 0.08024056255817413, + 0.08142638951539993, + 0.06290527433156967, + 0.5813673138618469 + ], + "credit_map": [ + 0.002989864209666848, + 0.002584246452897787, + 0.003342733019962907, + 0.0032722828909754753, + 0.002653108909726143, + 0.0028688900638371706, + 0.002997805131599307, + 0.0030660938937216997, + 0.0025798091664910316, + 0.0025960803031921387, + 0.00178951071575284, + 0.0015966264763846993, + 0.001910472521558404, + 0.002303245710209012, + 0.00138300028629601, + 0.0021182075142860413, + 0.004915677011013031, + 0.006003275979310274, + 0.002953536808490753, + 0.0016634323401376605, + 0.002608414040878415, + 0.002763709519058466, + 0.001187644200399518, + 0.001447691349312663, + 0.005498701240867376, + 0.0024991899263113737, + 0.003743560751900077, + 0.002534090541303158, + 0.013221447356045246, + 0.0038842479698359966, + 0.0024361221585422754, + 0.0041417707689106464, + 0.004598787985742092, + 0.0046074572019279, + 0.0049278163351118565, + 0.004172352608293295, + 0.009727433323860168, + 0.003047980135306716, + 0.003029034473001957, + 0.004628689959645271, + 0.006948056165128946, + 0.008652044460177422, + 0.004738274496048689, + 0.00290137086994946, + 0.0048406715504825115, + 0.008102267049252987, + 0.003462440799921751, + 0.014121265150606632, + 0.010831194929778576, + 0.01265969779342413, + 0.0067646438255906105, + 0.007673690561205149, + 0.009622039273381233, + 0.017013264819979668, + 0.006738313939422369, + 0.008937721140682697, + 0.011319541372358799, + 0.015589564107358456, + 0.01329975388944149, + 0.009607234038412571, + 0.008060772903263569, + 0.0072205024771392345, + 0.007766972295939922, + 0.008562052622437477, + 0.005968719720840454, + 0.011033587157726288, + 0.006500630639493465, + 0.013948909938335419, + 0.007075435481965542, + 0.005132783204317093, + 0.007303727325052023, + 0.0059414757415652275, + 0.0345175638794899, + 0.045566197484731674, + 0.0932830199599266, + 0.2115640789270401, + 0.04015989229083061, + 0.03501027822494507, + 0.06976494193077087, + 0.05150138959288597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.42887458205223083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle, wide shot captures a sleek, long-range glider navigating a sky that appears not as a natural expanse but as a pure Abstract Expressionist canvas. The backdrop is a violent, emotional maelstrom of brushstrokes, where thick, cold cadmium red, deep cobalt blue, and acid yellow clash against a chaotic field of stark black and white. The glider, with its polished, pearlescent white fuselage and incredibly slender wings, traverses this turbulent, painted space in a slow, deliberate roll. The camera tracks steadily along the aircraft as it rotates around its central axis with calm, almost balletic precision. As the glider flips, its glossy surface captures and reflects chaotic streaks of pure color, its wings temporarily painting the sky with ever-shifting, aggressive patterns. Just as the roll reaches its peak, the entire perspective is shattered by a sudden, violent jolt; the camera strikes with a raw, physical force that momentarily blurs the frame, disrupts the fluid movement, introduces a visceral shock, and tears through the scene's fragile aerial tranquility.", + "seed": 11680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_10/index_4995_seed_11680_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19781277822330595, + "vlm_logprob": -0.19781277822330595, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.19781277822330595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.633780837059021 + ], + "opposite_logprob_tokens": [ + -0.758780837059021 + ], + "logprob_reward": -0.633780837059021, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.633780837059021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2990581691265106, + 0.14933575689792633, + 0.05065510421991348, + 0.08025004714727402, + 0.04746546596288681, + 0.10324995219707489, + 0.026395833119750023, + 0.09149763733148575, + 0.11359408497810364, + 0.03849794343113899 + ], + "credit_map": [ + 0.01718904823064804, + 0.01751277968287468, + 0.024001849815249443, + 0.08540791273117065, + 0.08566664159297943, + 0.03190665692090988, + 0.019602444022893906, + 0.01777082122862339, + 0.0097312331199646, + 0.014368372969329357, + 0.015670733526349068, + 0.015491912141442299, + 0.04020138084888458, + 0.014346734620630741, + 0.014292538166046143, + 0.02523285150527954, + 0.003670457052066922, + 0.004556427244096994, + 0.008421243168413639, + 0.004842148628085852, + 0.004744415637105703, + 0.012719067744910717, + 0.005473593715578318, + 0.0062277535907924175, + 0.016915174201130867, + 0.007350110914558172, + 0.011112906970083714, + 0.010833607986569405, + 0.014497769996523857, + 0.00817652978003025, + 0.005048733204603195, + 0.006315214093774557, + 0.003645452903583646, + 0.014113002456724644, + 0.002567688003182411, + 0.003695691004395485, + 0.003690021811053157, + 0.0024657638277858496, + 0.01420935895293951, + 0.0030784867703914642, + 0.006501182448118925, + 0.005737381987273693, + 0.005333207082003355, + 0.00540569145232439, + 0.00990188866853714, + 0.026909785345196724, + 0.029348697513341904, + 0.014112120494246483, + 0.0028201949317008257, + 0.002653562230989337, + 0.0032325810752809048, + 0.002882937667891383, + 0.004294366575777531, + 0.0033978091087192297, + 0.004185325466096401, + 0.0029290569946169853, + 0.009362145327031612, + 0.00782268401235342, + 0.010650989599525928, + 0.010733325034379959, + 0.011389217339456081, + 0.013641118071973324, + 0.017846742644906044, + 0.010051416233181953, + 0.010630099102854729, + 0.015778301283717155, + 0.015389136038720608, + 0.015588438138365746, + 0.020887453109025955, + 0.011302308179438114, + 0.013157839886844158, + 0.010860507376492023, + 0.003088161814957857, + 0.008312752470374107, + 0.004211368504911661, + 0.005211965646594763, + 0.004397879354655743, + 0.003976528532803059, + 0.00354858604259789, + 0.005750702228397131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.633780837059021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266483407467604 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.005266483407467604, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266483407467604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41919389367103577, + 0.04667816311120987, + 0.11540348827838898, + 0.06692688167095184, + 0.1393618881702423, + 0.08838892728090286, + 0.03613751754164696, + 0.03200429305434227, + 0.033238302916288376, + 0.02266666851937771 + ], + "credit_map": [ + 0.017025824636220932, + 0.026056595146656036, + 0.022570623084902763, + 0.11303853243589401, + 0.1502159982919693, + 0.04840773344039917, + 0.024484412744641304, + 0.017394166439771652, + 0.0038390709087252617, + 0.00660698814317584, + 0.005379671696573496, + 0.0051421658135950565, + 0.00515670794993639, + 0.006775725167244673, + 0.006568507757037878, + 0.007209325674921274, + 0.014805824495851994, + 0.009486466646194458, + 0.029287315905094147, + 0.014096559956669807, + 0.011200848035514355, + 0.014333766885101795, + 0.010630637407302856, + 0.011562059633433819, + 0.01623360626399517, + 0.008740798570215702, + 0.014752295799553394, + 0.009331431239843369, + 0.004093215335160494, + 0.00605914369225502, + 0.004384942352771759, + 0.0033314432948827744, + 0.00989101268351078, + 0.04614241421222687, + 0.007079512346535921, + 0.006350718438625336, + 0.010469966568052769, + 0.007772245444357395, + 0.04274430871009827, + 0.008911692537367344, + 0.006616163998842239, + 0.006153847090899944, + 0.0052272058092057705, + 0.004534924402832985, + 0.01018559467047453, + 0.02044495940208435, + 0.02211572974920273, + 0.013110499829053879, + 0.004371838644146919, + 0.0032235917169600725, + 0.002987412502989173, + 0.003938867710530758, + 0.005770424380898476, + 0.0071307821199297905, + 0.004649512004107237, + 0.004065085668116808, + 0.0034861592575907707, + 0.003077901666983962, + 0.004239231813699007, + 0.0036414810456335545, + 0.003698345273733139, + 0.0047528729774057865, + 0.0055526914075016975, + 0.0035556082148104906, + 0.00329818413592875, + 0.004465980920940638, + 0.003833228489384055, + 0.004610524978488684, + 0.0056691556237638, + 0.003261803649365902, + 0.0046303775161504745, + 0.0034690427128225565, + 0.001781669445335865, + 0.0021885191090404987, + 0.0021668034605681896, + 0.0018889853963628411, + 0.0015768638113513589, + 0.0017235423438251019, + 0.004120055586099625, + 0.007220226339995861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.005266483407467604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989104986190796 + ], + "opposite_logprob_tokens": [ + -1.8048911094665527 + ], + "logprob_reward": -0.17989104986190796, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989104986190796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.302163690328598, + 0.09718464314937592, + 0.09763973206281662, + 0.13275791704654694, + 0.05903434008359909, + 0.0673409029841423, + 0.03955559432506561, + 0.03613097965717316, + 0.10238255560398102, + 0.06580974161624908 + ], + "credit_map": [ + 0.015898462384939194, + 0.017454473301768303, + 0.02092791348695755, + 0.07552682608366013, + 0.10406892746686935, + 0.034366361796855927, + 0.018660591915249825, + 0.015260107815265656, + 0.006923002190887928, + 0.008960224688053131, + 0.011989647522568703, + 0.01156835537403822, + 0.015961460769176483, + 0.017784496769309044, + 0.010587689466774464, + 0.01340976357460022, + 0.011132187210023403, + 0.009067436680197716, + 0.014117290265858173, + 0.010322876274585724, + 0.013018200173974037, + 0.015820300206542015, + 0.009983891621232033, + 0.014177544042468071, + 0.028780154883861542, + 0.01142360083758831, + 0.021482275798916817, + 0.0190243162214756, + 0.02125714160501957, + 0.012615159153938293, + 0.009397899731993675, + 0.008777376264333725, + 0.003923241514712572, + 0.018618177622556686, + 0.0031752213835716248, + 0.00349529180675745, + 0.003980185370892286, + 0.003389887511730194, + 0.018695147708058357, + 0.0037571885623037815, + 0.006958822254091501, + 0.005576067604124546, + 0.0036145527847111225, + 0.003685787320137024, + 0.009277449920773506, + 0.015413723886013031, + 0.012375078164041042, + 0.010439419187605381, + 0.004570172633975744, + 0.004399067256599665, + 0.002952049020677805, + 0.0034333260264247656, + 0.007506418041884899, + 0.006685574539005756, + 0.005536598619073629, + 0.004472389817237854, + 0.004349606577306986, + 0.00444593233987689, + 0.0031789455097168684, + 0.0036559670697897673, + 0.002998167183250189, + 0.0067590633407235146, + 0.004758258815854788, + 0.0059850383549928665, + 0.009940708987414837, + 0.013370491564273834, + 0.014574195258319378, + 0.01444946601986885, + 0.015018977224826813, + 0.010635084472596645, + 0.012152389623224735, + 0.01224124152213335, + 0.004424410406500101, + 0.01398053765296936, + 0.005162513814866543, + 0.003978068474680185, + 0.005298287142068148, + 0.005544052459299564, + 0.0076306965202093124, + 0.01979117840528488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.17989104986190796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785539470613003 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785539470613003, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785539470613003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33010563254356384, + 0.21915553510189056, + 0.06524282693862915, + 0.06864257156848907, + 0.04698578640818596, + 0.037117138504981995, + 0.0584232471883297, + 0.0555243194103241, + 0.0645240843296051, + 0.05427886173129082 + ], + "credit_map": [ + 0.014071165584027767, + 0.019959457218647003, + 0.020823055878281593, + 0.09264642000198364, + 0.10960493981838226, + 0.03711171820759773, + 0.01913517341017723, + 0.016753708943724632, + 0.018580172210931778, + 0.019119158387184143, + 0.03499431908130646, + 0.01899896375834942, + 0.051125723868608475, + 0.02495618537068367, + 0.018251309171319008, + 0.03312971442937851, + 0.006544038653373718, + 0.008135700598359108, + 0.009898902848362923, + 0.010588770732283592, + 0.006592740770429373, + 0.007741191890090704, + 0.007191441487520933, + 0.008550038561224937, + 0.012286721728742123, + 0.00859729666262865, + 0.00933280773460865, + 0.017142029479146004, + 0.006446505431085825, + 0.004278723616153002, + 0.005394921638071537, + 0.005163560155779123, + 0.002780655864626169, + 0.01323795598000288, + 0.004316464997828007, + 0.005461648106575012, + 0.002853651065379381, + 0.002118088770657778, + 0.013265417888760567, + 0.002951904432848096, + 0.003690658835694194, + 0.004068358335644007, + 0.0031572128646075726, + 0.0037289364263415337, + 0.004692134913057089, + 0.006633225828409195, + 0.0067749302834272385, + 0.004371682181954384, + 0.006180816795676947, + 0.005966662894934416, + 0.005987692158669233, + 0.007482727058231831, + 0.010357464663684368, + 0.008211402222514153, + 0.008052960969507694, + 0.006183525547385216, + 0.007196828722953796, + 0.005703565198928118, + 0.006033142562955618, + 0.0051267389208078384, + 0.005213046912103891, + 0.010061598382890224, + 0.008294505067169666, + 0.007894893176853657, + 0.00484599219635129, + 0.008627627976238728, + 0.007955251261591911, + 0.009859268553555012, + 0.010505220852792263, + 0.006158831994980574, + 0.008718500845134258, + 0.007853393442928791, + 0.004027053248137236, + 0.012882629409432411, + 0.00453444616869092, + 0.003951850347220898, + 0.005833157338202, + 0.005232048220932484, + 0.004521212074905634, + 0.013296463526785374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009785539470613003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek glider with a polished, pearlescent white fuselage and incredibly slender wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a chaotic Abstract Expressionist canvas featuring thick brushstrokes of cadmium red, cobalt blue, acid yellow, black, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the glider perform a slow, deliberate roll while traversing the painted space?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track steadily along the aircraft as it rotates, capturing reflections of color on its glossy surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with a sudden, violent jolt that blurs the frame and disrupts the smooth movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033998131752014 + ], + "opposite_logprob_tokens": [ + -1.9103399515151978 + ], + "logprob_reward": -0.16033998131752014, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033998131752014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23557858169078827, + 0.11825679242610931, + 0.08935025334358215, + 0.07848811894655228, + 0.021716441959142685, + 0.1918342113494873, + 0.03610704839229584, + 0.033733684569597244, + 0.04038271680474281, + 0.15455220639705658 + ], + "credit_map": [ + 0.01067422516644001, + 0.0138924028724432, + 0.01397081557661295, + 0.06544896960258484, + 0.08408179879188538, + 0.026171699166297913, + 0.012883065268397331, + 0.008455569855868816, + 0.008942067623138428, + 0.01278137881308794, + 0.015187946148216724, + 0.010567393153905869, + 0.02830292098224163, + 0.011221883818507195, + 0.01162955816835165, + 0.019623637199401855, + 0.00832324754446745, + 0.00844300165772438, + 0.01712571270763874, + 0.012775720097124577, + 0.007464960683137178, + 0.013419373892247677, + 0.007840300910174847, + 0.013957925140857697, + 0.009023836813867092, + 0.006512731779366732, + 0.017613189294934273, + 0.007936879061162472, + 0.016915831714868546, + 0.009742217138409615, + 0.004020493011921644, + 0.006722933147102594, + 0.002855866216123104, + 0.0033697772305458784, + 0.0032037158962339163, + 0.0036257714964449406, + 0.0029614483937621117, + 0.0018924114992842078, + 0.0016998294740915298, + 0.002107620006427169, + 0.012164875864982605, + 0.011275517754256725, + 0.008956470526754856, + 0.008350626565515995, + 0.018726862967014313, + 0.04856102913618088, + 0.05605364963412285, + 0.02774515189230442, + 0.0073360176756978035, + 0.005632823798805475, + 0.0027857813984155655, + 0.003559516277164221, + 0.004296082071959972, + 0.003918920643627644, + 0.003829379566013813, + 0.004748527426272631, + 0.004579026717692614, + 0.003496853867545724, + 0.0025968453846871853, + 0.0025743748992681503, + 0.003330841427668929, + 0.007179519161581993, + 0.0035786305088549852, + 0.006397589575499296, + 0.0044688452035188675, + 0.004220736213028431, + 0.0049449848011136055, + 0.004906477872282267, + 0.006224945653229952, + 0.005415157414972782, + 0.005571532063186169, + 0.0046300324611365795, + 0.01595127210021019, + 0.0394047275185585, + 0.01061832532286644, + 0.011215274222195148, + 0.02133500762283802, + 0.016936343163251877, + 0.013322403654456139, + 0.025768835097551346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.16033998131752014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_tensor([2273, 2273])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_tensor([2273, 2273])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..57cd37ed5b8825e4a36500a4555d1b58db70ddb9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_tensor([2273, 2273])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot, utilizing an incredibly shallow depth of field, focuses with razor-sharp precision on the vibrant, cadmium-red tip of an oil stick. The stick's waxy, creamy texture is visible in minute details, slightly deformed and softened by the application of pressure. As it is vigorously spun and dragged across the rough white canvas, the camera captures the dynamic action in microscopic slow motion. The stick leaves a thick, viscous trail across the canvas while, with controlled violence, flinging countless tiny particles of pure, saturated pigment into the air. These particles form a chaotic yet beautiful spiral, suspended briefly like a colorful galaxy. Bright, warm studio light, emanating from the side, shimmers on the smooth, damp pigment and the stick's waxy surface, while the background, including the artist's unseen hand and other supplies, completely dissolves into a soft, painterly bokeh of warm, blurred tones, creating a powerful, abstract atmosphere. The image reflects intense, focused creative energy.", + "seed": 8957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_2273_seed_8957_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3640553455799818, + "vlm_logprob": -0.3640553455799818, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.3640553455799818, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017829034477472305 + ], + "opposite_logprob_tokens": [ + -4.267828941345215 + ], + "logprob_reward": -0.017829034477472305, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017829034477472305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049559395760297775, + 0.01279436331242323, + 0.03971201553940773, + 0.010680737905204296, + 0.12459666281938553, + 0.014444013126194477, + 0.0060040936805307865, + 0.039427805691957474, + 0.653845489025116, + 0.04893546923995018 + ], + "credit_map": [ + 0.004510586615651846, + 0.006592805963009596, + 0.00561709376052022, + 0.009944447316229343, + 0.006407381966710091, + 0.005017498042434454, + 0.005716516170650721, + 0.005753069184720516, + 0.0017914441414177418, + 0.0013863366330042481, + 0.0016023784410208464, + 0.0015429258346557617, + 0.0014725147048011422, + 0.0012961095198988914, + 0.001752599491737783, + 0.0019500558264553547, + 0.00298356544226408, + 0.0035530938766896725, + 0.004341743420809507, + 0.005349291954189539, + 0.005353336688131094, + 0.006430187728255987, + 0.006536619737744331, + 0.005164174363017082, + 0.001168898306787014, + 0.0015503503382205963, + 0.0010270696366205812, + 0.001507416251115501, + 0.0015302179381251335, + 0.0013164585689082742, + 0.0011424830881878734, + 0.0014378431951627135, + 0.020507514476776123, + 0.020941205322742462, + 0.033754970878362656, + 0.006601390894502401, + 0.012395970523357391, + 0.018607810139656067, + 0.006084483116865158, + 0.005703320726752281, + 0.0016396351857110858, + 0.00125503062736243, + 0.002667873864993453, + 0.001064845360815525, + 0.001319651142694056, + 0.002371129347011447, + 0.001221898477524519, + 0.002903949934989214, + 0.00075049779843539, + 0.0007771822274662554, + 0.0009928622748702765, + 0.0007570515153929591, + 0.0008156524272635579, + 0.0008261333568952978, + 0.0004690663772635162, + 0.0006156476447358727, + 0.0029252981767058372, + 0.008253938518464565, + 0.016127197071909904, + 0.0013754266547039151, + 0.002273795660585165, + 0.005716707557439804, + 0.0011746742529794574, + 0.001580767217092216, + 0.0647289976477623, + 0.12953804433345795, + 0.2828049957752228, + 0.02437971532344818, + 0.06614331901073456, + 0.03341313451528549, + 0.02485695295035839, + 0.027980294078588486, + 0.007883592508733273, + 0.017964430153369904, + 0.003697889856994152, + 0.00277652102522552, + 0.008232932537794113, + 0.0035120651591569185, + 0.0023665374610573053, + 0.0025014998391270638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.017829034477472305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018231434747576714 + ], + "opposite_logprob_tokens": [ + -4.018231391906738 + ], + "logprob_reward": -0.018231434747576714, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018231434747576714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07064526528120041, + 0.023099280893802643, + 0.09382446110248566, + 0.02611730434000492, + 0.1757865995168686, + 0.02353121154010296, + 0.010908057913184166, + 0.1288832724094391, + 0.3185003697872162, + 0.12870420515537262 + ], + "credit_map": [ + 0.007430781610310078, + 0.011992218904197216, + 0.008545199409127235, + 0.011278713122010231, + 0.007262948900461197, + 0.006393905263394117, + 0.008868010714650154, + 0.008873485028743744, + 0.0034040517639368773, + 0.002555177081376314, + 0.003047178266569972, + 0.002702338621020317, + 0.002882755594328046, + 0.002001416403800249, + 0.002323675900697708, + 0.004182684700936079, + 0.007051336579024792, + 0.009979886934161186, + 0.010538412258028984, + 0.014881829731166363, + 0.0130719393491745, + 0.009630776941776276, + 0.016053274273872375, + 0.012617001309990883, + 0.0024098525755107403, + 0.004485254641622305, + 0.002391988644376397, + 0.004011656157672405, + 0.003427779534831643, + 0.0036070647183805704, + 0.0025438000448048115, + 0.003239907557144761, + 0.026093291118741035, + 0.03719470277428627, + 0.06067412346601486, + 0.010252241976559162, + 0.011541804298758507, + 0.016417421400547028, + 0.006414677947759628, + 0.007198338396847248, + 0.002497519599273801, + 0.001998841529712081, + 0.0044503468088805676, + 0.0020555865485221148, + 0.002154039219021797, + 0.002877514809370041, + 0.0025194543413817883, + 0.00497790751978755, + 0.0013078724732622504, + 0.0016620203386992216, + 0.0015192587161436677, + 0.0012827053433284163, + 0.0014314079890027642, + 0.0017028843285515904, + 0.0008825630065985024, + 0.0011193460086360574, + 0.010506942868232727, + 0.025550127029418945, + 0.05305822566151619, + 0.004491402767598629, + 0.008458326570689678, + 0.017508888617157936, + 0.004180707037448883, + 0.005128643941134214, + 0.027396569028496742, + 0.059653881937265396, + 0.13084463775157928, + 0.013117421418428421, + 0.030091630294919014, + 0.017906630411744118, + 0.016948560252785683, + 0.02254101075232029, + 0.02263636887073517, + 0.036350447684526443, + 0.01915574260056019, + 0.006627473048865795, + 0.022817866876721382, + 0.009112243540585041, + 0.005814439617097378, + 0.0061896261759102345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.018231434747576714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.701557993888855 + ], + "opposite_logprob_tokens": [ + -0.2015579640865326 + ], + "logprob_reward": -1.701557993888855, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.701557993888855, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04541189596056938, + 0.018252786248922348, + 0.07045100629329681, + 0.07914664596319199, + 0.24268795549869537, + 0.03819754719734192, + 0.006167404353618622, + 0.08326123654842377, + 0.27214235067367554, + 0.14428120851516724 + ], + "credit_map": [ + 0.004494875203818083, + 0.004922081250697374, + 0.005280972924083471, + 0.009502027183771133, + 0.00438675144687295, + 0.00567179499194026, + 0.005664025899022818, + 0.005489361006766558, + 0.0025874748826026917, + 0.0021584592759609222, + 0.0023591758217662573, + 0.0023305874783545732, + 0.0015564034692943096, + 0.0020089480094611645, + 0.002597420709207654, + 0.002654315670952201, + 0.008257880806922913, + 0.0090809166431427, + 0.01084381528198719, + 0.00843860488384962, + 0.009410298429429531, + 0.007858041673898697, + 0.007999508641660213, + 0.008561928756535053, + 0.008463776670396328, + 0.013647087849676609, + 0.007402391172945499, + 0.011705607175827026, + 0.010209940373897552, + 0.010570553131401539, + 0.007945922203361988, + 0.009201355278491974, + 0.03810795024037361, + 0.04851263388991356, + 0.08469724655151367, + 0.014450629241764545, + 0.01481990423053503, + 0.02178717963397503, + 0.00921808835119009, + 0.01109430380165577, + 0.004316127859055996, + 0.003466973779723048, + 0.007787174079567194, + 0.003974009305238724, + 0.0035880121868103743, + 0.005368810147047043, + 0.004080601967871189, + 0.005615835078060627, + 0.0006793473730795085, + 0.000727747508790344, + 0.0011528537143021822, + 0.0008084058645181358, + 0.0007844445062801242, + 0.0007128330762498081, + 0.0006628534756600857, + 0.0006389179034158587, + 0.006190858315676451, + 0.017040861770510674, + 0.033929042518138885, + 0.0032568841706961393, + 0.005213819444179535, + 0.011339962482452393, + 0.0025227232836186886, + 0.0037670766469091177, + 0.029330721125006676, + 0.05177959427237511, + 0.11632312089204788, + 0.010504700243473053, + 0.02836916781961918, + 0.013057801872491837, + 0.010579383932054043, + 0.012197846546769142, + 0.028291966766119003, + 0.04002893716096878, + 0.019504105672240257, + 0.007009897846728563, + 0.02430523931980133, + 0.010671793483197689, + 0.006788729224354029, + 0.007680525071918964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -1.701557993888855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903824001550674 + ], + "opposite_logprob_tokens": [ + -2.579038143157959 + ], + "logprob_reward": -0.07903824001550674, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903824001550674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02395513840019703, + 0.025078510865569115, + 0.07951073348522186, + 0.04852413013577461, + 0.4554589092731476, + 0.0214825551956892, + 0.007556576281785965, + 0.11818395555019379, + 0.10075264424085617, + 0.11949680000543594 + ], + "credit_map": [ + 0.00322711025364697, + 0.002297402825206518, + 0.0027570342645049095, + 0.004401249345391989, + 0.0022605261765420437, + 0.002817797474563122, + 0.003513361094519496, + 0.0026806569658219814, + 0.002801687689498067, + 0.003321200842037797, + 0.003791365772485733, + 0.003031702246516943, + 0.0027274389285594225, + 0.0024663766380399466, + 0.0035434330347925425, + 0.003395308507606387, + 0.009629298932850361, + 0.014741840772330761, + 0.012048721313476562, + 0.008503763936460018, + 0.009237566031515598, + 0.006310540717095137, + 0.005742376204580069, + 0.013296625576913357, + 0.005550051108002663, + 0.008357174694538116, + 0.004136458970606327, + 0.007124129217118025, + 0.0061385068111121655, + 0.0064252568408846855, + 0.004970175679773092, + 0.005822375882416964, + 0.08127011358737946, + 0.0966348871588707, + 0.162748783826828, + 0.023847047239542007, + 0.023804426193237305, + 0.039956845343112946, + 0.01220832858234644, + 0.014988506212830544, + 0.002453880850225687, + 0.002016110345721245, + 0.004168321844190359, + 0.002463787328451872, + 0.0022446662187576294, + 0.002590779447928071, + 0.002253409009426832, + 0.0032915989868342876, + 0.0009610409033484757, + 0.0008094900986179709, + 0.0012026219628751278, + 0.0007367135840468109, + 0.000758267764467746, + 0.0016013924032449722, + 0.0007158046355471015, + 0.0007712451042607427, + 0.00888095423579216, + 0.02439378760755062, + 0.048992667347192764, + 0.0040235985070466995, + 0.007338446564972401, + 0.016803033649921417, + 0.003322951029986143, + 0.004428519401699305, + 0.012398237362504005, + 0.018581749871373177, + 0.04212602972984314, + 0.00450441800057888, + 0.010752025991678238, + 0.0048459735698997974, + 0.003657953580841422, + 0.0038862561341375113, + 0.022570153698325157, + 0.04003879427909851, + 0.009477670304477215, + 0.00586409866809845, + 0.020820986479520798, + 0.009127714671194553, + 0.0056876554153859615, + 0.005909726954996586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.07903824001550674 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036200247704982758 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036200247704982758, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036200247704982758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06903227418661118, + 0.01713879592716694, + 0.07819893956184387, + 0.05687669664621353, + 0.1646304875612259, + 0.0410580150783062, + 0.021992230787873268, + 0.12391765415668488, + 0.34441444277763367, + 0.08274044096469879 + ], + "credit_map": [ + 0.007928273640573025, + 0.010515475645661354, + 0.008098779246211052, + 0.01174046192318201, + 0.007670429069548845, + 0.009744352661073208, + 0.006806490011513233, + 0.006528009194880724, + 0.0029431134462356567, + 0.001863753772340715, + 0.002199060982093215, + 0.0022673949133604765, + 0.0014907456934452057, + 0.0016296985559165478, + 0.0019341715378686786, + 0.002810858190059662, + 0.009093420580029488, + 0.00908094272017479, + 0.011303824372589588, + 0.01017431728541851, + 0.00879979319870472, + 0.009444219060242176, + 0.009856697171926498, + 0.010445726104080677, + 0.006460231728851795, + 0.008516775444149971, + 0.007202426437288523, + 0.008865075185894966, + 0.0072791678830981255, + 0.006482842843979597, + 0.005059293936938047, + 0.007010878995060921, + 0.0253804549574852, + 0.031422533094882965, + 0.05483322590589523, + 0.01052145753055811, + 0.012313694693148136, + 0.016126878559589386, + 0.00670525012537837, + 0.007326989434659481, + 0.004572893492877483, + 0.004074846860021353, + 0.008385209366679192, + 0.004120721947401762, + 0.003971523605287075, + 0.00490646343678236, + 0.004711742978543043, + 0.006314614322036505, + 0.0029960470274090767, + 0.0025472857523709536, + 0.0033970214426517487, + 0.002533896127715707, + 0.0026141053531318903, + 0.0036557940766215324, + 0.001883766264654696, + 0.002364315325394273, + 0.010199285112321377, + 0.02534475363790989, + 0.05166753754019737, + 0.004385281354188919, + 0.008203038945794106, + 0.016871636733412743, + 0.0034834002144634724, + 0.003762732958421111, + 0.03925960510969162, + 0.06126583740115166, + 0.14334799349308014, + 0.013545825146138668, + 0.03680393472313881, + 0.015875661745667458, + 0.01649613305926323, + 0.0178194809705019, + 0.017117798328399658, + 0.009531357325613499, + 0.020072851330041885, + 0.007359911687672138, + 0.010123697109520435, + 0.005638810805976391, + 0.0050380416214466095, + 0.0078579718247056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.0036200247704982758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, utilizing an incredibly shallow depth of field, focuses with razor-sharp precision on the vibrant, cadmium-red tip of an oil stick. The stick's waxy, creamy texture is visible in minute details, slightly deformed and softened by the application of pressure. As it is vigorously spun and dragged across the rough white canvas, the camera captures the dynamic action in microscopic slow motion. The stick leaves a thick, viscous trail across the canvas while, with controlled violence, flinging countless tiny particles of pure, saturated pigment into the air. These particles form a chaotic yet beautiful spiral, suspended briefly like a colorful galaxy. Bright, warm studio light, emanating from the side, shimmers on the smooth, damp pigment and the stick's waxy surface, while the background, including the artist's unseen hand and other supplies, completely dissolves into a soft, painterly bokeh of warm, blurred tones, creating a powerful, abstract atmosphere. The image reflects intense, focused creative energy.", + "seed": 8958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_2273_seed_8958_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3817281281575561, + "vlm_logprob": -0.3817281281575561, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.3817281281575561, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0350768007338047 + ], + "opposite_logprob_tokens": [ + -3.535076856613159 + ], + "logprob_reward": -0.0350768007338047, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0350768007338047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08425890654325485, + 0.043406225740909576, + 0.10457195341587067, + 0.059327516704797745, + 0.060697950422763824, + 0.14599180221557617, + 0.2503628432750702, + 0.11625469475984573, + 0.09814094752073288, + 0.03698708862066269 + ], + "credit_map": [ + 0.018016889691352844, + 0.017923397943377495, + 0.0030790145974606276, + 0.0050918906927108765, + 0.00989394448697567, + 0.00976548157632351, + 0.0035622429568320513, + 0.016926046460866928, + 0.005643290467560291, + 0.005077779293060303, + 0.0022186238784343004, + 0.002887879265472293, + 0.010610368102788925, + 0.0072404490783810616, + 0.004462535493075848, + 0.005265301093459129, + 0.014659125357866287, + 0.009073053486645222, + 0.006981987040489912, + 0.018982812762260437, + 0.022884920239448547, + 0.01224543247371912, + 0.008885920979082584, + 0.010858704335987568, + 0.007660270668566227, + 0.005978187546133995, + 0.004637439735233784, + 0.003823622828349471, + 0.007614972535520792, + 0.005497467704117298, + 0.00833167415112257, + 0.01578388549387455, + 0.007321606855839491, + 0.00946778804063797, + 0.005285024177283049, + 0.008264024741947651, + 0.008361978456377983, + 0.006217661779373884, + 0.006774227134883404, + 0.009005647152662277, + 0.010835603810846806, + 0.020785359665751457, + 0.013781912624835968, + 0.0111431535333395, + 0.030086373910307884, + 0.02803529053926468, + 0.015044474974274635, + 0.01627965085208416, + 0.06095945090055466, + 0.018656043335795403, + 0.021070046350359917, + 0.012453591451048851, + 0.023977184668183327, + 0.06623712927103043, + 0.013508607633411884, + 0.03350081294775009, + 0.012250960804522038, + 0.011345521546900272, + 0.022266631945967674, + 0.01920464262366295, + 0.013571514748036861, + 0.012053017504513264, + 0.012499955482780933, + 0.013062460348010063, + 0.018241997808218002, + 0.005678767804056406, + 0.020398231223225594, + 0.009804018773138523, + 0.008661872707307339, + 0.011529305949807167, + 0.014856155030429363, + 0.008970609866082668, + 0.003266570158302784, + 0.0037799589335918427, + 0.004405528772622347, + 0.004786069039255381, + 0.0062389797531068325, + 0.004462165758013725, + 0.00570959085598588, + 0.0043382300063967705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0350768007338047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033774398267269135 + ], + "opposite_logprob_tokens": [ + -3.4087743759155273 + ], + "logprob_reward": -0.033774398267269135, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033774398267269135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0631018802523613, + 0.07404033094644547, + 0.09427172690629959, + 0.1411973387002945, + 0.1066654846072197, + 0.06357012689113617, + 0.26499953866004944, + 0.10761836171150208, + 0.05796528980135918, + 0.026569906622171402 + ], + "credit_map": [ + 0.007834167219698429, + 0.012088827788829803, + 0.00434334110468626, + 0.007889864966273308, + 0.008780671283602715, + 0.007034897338598967, + 0.0076813665218651295, + 0.007448745891451836, + 0.010600818321108818, + 0.009290974587202072, + 0.004800484515726566, + 0.008139916695654392, + 0.014800620265305042, + 0.01067309733480215, + 0.008050084114074707, + 0.007684339303523302, + 0.011934511363506317, + 0.008232471533119678, + 0.0068445331417024136, + 0.01565687172114849, + 0.018007870763540268, + 0.010860541835427284, + 0.009348687715828419, + 0.013386240229010582, + 0.019863514229655266, + 0.01546716969460249, + 0.011526999995112419, + 0.01036166027188301, + 0.014210579916834831, + 0.016409270465373993, + 0.014620521105825901, + 0.03873762860894203, + 0.01303136721253395, + 0.012842012569308281, + 0.009410770609974861, + 0.01819598861038685, + 0.012609264813363552, + 0.011970072984695435, + 0.011481444351375103, + 0.017124559730291367, + 0.006069630850106478, + 0.01035934966057539, + 0.005419979803264141, + 0.005985429976135492, + 0.011903713457286358, + 0.009460106492042542, + 0.007245955523103476, + 0.0071259657852351665, + 0.0782981663942337, + 0.018282687291502953, + 0.016321882605552673, + 0.0118949543684721, + 0.020536383613944054, + 0.05503038316965103, + 0.013956816866993904, + 0.05067828297615051, + 0.00887159164994955, + 0.016819283366203308, + 0.012858934700489044, + 0.018403710797429085, + 0.012818219140172005, + 0.012259604409337044, + 0.013930276036262512, + 0.01165674440562725, + 0.009020359255373478, + 0.003571789711713791, + 0.012107261456549168, + 0.007287709508091211, + 0.005369770340621471, + 0.004218395333737135, + 0.011208564043045044, + 0.0051814415492117405, + 0.0026719102170318365, + 0.002621763153001666, + 0.0032163369469344616, + 0.003749387338757515, + 0.005159834865480661, + 0.003402940696105361, + 0.00356866791844368, + 0.002179065253585577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.033774398267269135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015554904937744 + ], + "opposite_logprob_tokens": [ + -0.20155543088912964 + ], + "logprob_reward": -1.7015554904937744, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015554904937744, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13069134950637817, + 0.06570058315992355, + 0.032811425626277924, + 0.0974464863538742, + 0.03115461952984333, + 0.13319958746433258, + 0.33523985743522644, + 0.1084911897778511, + 0.05173219367861748, + 0.013532704673707485 + ], + "credit_map": [ + 0.05364188179373741, + 0.008953865617513657, + 0.0032412901055067778, + 0.0035761899780482054, + 0.005035030655562878, + 0.007111272308975458, + 0.004061294253915548, + 0.045070528984069824, + 0.009700574912130833, + 0.009811020456254482, + 0.0028879919555038214, + 0.004966061562299728, + 0.013392427936196327, + 0.012023922987282276, + 0.0065381815657019615, + 0.006380401086062193, + 0.004277065861970186, + 0.002883120672777295, + 0.0028266103472560644, + 0.004821405746042728, + 0.0038474551402032375, + 0.005547856912016869, + 0.004325975198298693, + 0.004281936679035425, + 0.013662870042026043, + 0.01052556186914444, + 0.008504711091518402, + 0.008944074623286724, + 0.012774952687323093, + 0.010100085288286209, + 0.006967047695070505, + 0.02596719190478325, + 0.0029477220959961414, + 0.0036388556472957134, + 0.002872258657589555, + 0.005974202416837215, + 0.0043889801017940044, + 0.0037132399156689644, + 0.0034730348270386457, + 0.0041463240049779415, + 0.009748074226081371, + 0.02420641854405403, + 0.009355264715850353, + 0.011047865264117718, + 0.029133684933185577, + 0.0209487434476614, + 0.01656150259077549, + 0.012198028154671192, + 0.08419906347990036, + 0.025196054950356483, + 0.03765973076224327, + 0.01696796528995037, + 0.02258557453751564, + 0.028779858723282814, + 0.028179848566651344, + 0.09167178720235825, + 0.008758177049458027, + 0.015904607251286507, + 0.01387201901525259, + 0.013504458591341972, + 0.011182188987731934, + 0.012706988491117954, + 0.018161414191126823, + 0.014401334337890148, + 0.008208485320210457, + 0.004433031193912029, + 0.01030874066054821, + 0.007455947808921337, + 0.0038398858159780502, + 0.0055684796534478664, + 0.00754647608846426, + 0.004371145274490118, + 0.0012956710997968912, + 0.0014464444248005748, + 0.0016799927689135075, + 0.002268855692818761, + 0.001961377216503024, + 0.0011448939330875874, + 0.0020760190673172474, + 0.0016594494227319956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -1.7015554904937744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12714600563049316 + ], + "opposite_logprob_tokens": [ + -2.127146005630493 + ], + "logprob_reward": -0.12714600563049316, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12714600563049316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14387862384319305, + 0.04286660999059677, + 0.0277460478246212, + 0.07677026838064194, + 0.04051857069134712, + 0.17282606661319733, + 0.3259701728820801, + 0.07987399399280548, + 0.06588269770145416, + 0.023666875436902046 + ], + "credit_map": [ + 0.0590418204665184, + 0.009663473814725876, + 0.0035503010731190443, + 0.0037753542419523, + 0.0054410905577242374, + 0.00789918564260006, + 0.004342015367001295, + 0.05016537383198738, + 0.006259013898670673, + 0.0061040339060127735, + 0.0020812826696783304, + 0.0033646889496594667, + 0.009067482315003872, + 0.007265049498528242, + 0.004478786140680313, + 0.004246277268975973, + 0.0033866253215819597, + 0.0025426519569009542, + 0.0018753694603219628, + 0.004642867017537355, + 0.0053778099827468395, + 0.003684057854115963, + 0.002702706726267934, + 0.003533960087224841, + 0.010739672929048538, + 0.009073380380868912, + 0.006795146968215704, + 0.005949400365352631, + 0.008580363355576992, + 0.009186448529362679, + 0.006117455195635557, + 0.020328396931290627, + 0.003570982487872243, + 0.00573984207585454, + 0.004359705373644829, + 0.007649354636669159, + 0.005945401731878519, + 0.004501641262322664, + 0.004551802761852741, + 0.0041998401284217834, + 0.012070725671947002, + 0.029917538166046143, + 0.013505379669368267, + 0.01426932867616415, + 0.037785857915878296, + 0.028744250535964966, + 0.01984810084104538, + 0.016684873029589653, + 0.0953059121966362, + 0.02285854145884514, + 0.020966140553355217, + 0.01503114216029644, + 0.022051937878131866, + 0.06397154182195663, + 0.019018154591321945, + 0.06676684319972992, + 0.0061545539647340775, + 0.012111508287489414, + 0.010402624495327473, + 0.01045992411673069, + 0.008913228288292885, + 0.009620429947972298, + 0.012428968213498592, + 0.009782752022147179, + 0.01187409833073616, + 0.004767716396600008, + 0.014343501068651676, + 0.009580659680068493, + 0.0045461575500667095, + 0.0063863713294267654, + 0.008584310300648212, + 0.005799883510917425, + 0.0018523026956245303, + 0.0021564115304499865, + 0.003265024395659566, + 0.004361413884907961, + 0.0026810721028596163, + 0.0017684124177321792, + 0.004549461882561445, + 0.0030327741988003254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.12714600563049316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108794566243887 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.01108794566243887, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108794566243887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1006748229265213, + 0.04718845710158348, + 0.054739341139793396, + 0.10771618038415909, + 0.029568659141659737, + 0.07896401733160019, + 0.32407546043395996, + 0.1565731018781662, + 0.07809168845415115, + 0.02240827865898609 + ], + "credit_map": [ + 0.037192776799201965, + 0.007015237584710121, + 0.003163217334076762, + 0.005383202340453863, + 0.005750758107751608, + 0.0061170794069767, + 0.004863937851041555, + 0.031188614666461945, + 0.0065194591879844666, + 0.005935786757618189, + 0.0022283424623310566, + 0.0038944941479712725, + 0.009602537378668785, + 0.008614284917712212, + 0.005778070539236069, + 0.004615478683263063, + 0.00653541786596179, + 0.004792551510035992, + 0.00436597503721714, + 0.007840464822947979, + 0.0104298060759902, + 0.00897193793207407, + 0.0060733770951628685, + 0.005729815922677517, + 0.01439383439719677, + 0.010596057400107384, + 0.009131655097007751, + 0.01013626717031002, + 0.01398750115185976, + 0.0108304088935256, + 0.00848631002008915, + 0.0301541518419981, + 0.0025191318709403276, + 0.004346319939941168, + 0.0027535767294466496, + 0.005059551913291216, + 0.004896694328635931, + 0.0036334171891212463, + 0.0031193336471915245, + 0.00324063072912395, + 0.0068616969510912895, + 0.013363192789256573, + 0.005505991633981466, + 0.007130068726837635, + 0.016277845948934555, + 0.012072279118001461, + 0.008248935453593731, + 0.009504006244242191, + 0.07496561110019684, + 0.023323649540543556, + 0.036502640694379807, + 0.016781460493803024, + 0.022564435377717018, + 0.03343849629163742, + 0.027969785034656525, + 0.08852940052747726, + 0.013080283999443054, + 0.02536637894809246, + 0.018769308924674988, + 0.023650676012039185, + 0.017951061949133873, + 0.016936127096414566, + 0.02219308540225029, + 0.018626181408762932, + 0.0073534660041332245, + 0.0062718819826841354, + 0.011108068749308586, + 0.008541201241314411, + 0.012000806629657745, + 0.011503707617521286, + 0.011937553994357586, + 0.009375002235174179, + 0.0023861576337367296, + 0.0023245597258210182, + 0.0033436415251344442, + 0.0031782647129148245, + 0.0028378013521432877, + 0.0018861927092075348, + 0.0037603506352752447, + 0.002691308967769146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.01108794566243887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, utilizing an incredibly shallow depth of field, focuses with razor-sharp precision on the vibrant, cadmium-red tip of an oil stick. The stick's waxy, creamy texture is visible in minute details, slightly deformed and softened by the application of pressure. As it is vigorously spun and dragged across the rough white canvas, the camera captures the dynamic action in microscopic slow motion. The stick leaves a thick, viscous trail across the canvas while, with controlled violence, flinging countless tiny particles of pure, saturated pigment into the air. These particles form a chaotic yet beautiful spiral, suspended briefly like a colorful galaxy. Bright, warm studio light, emanating from the side, shimmers on the smooth, damp pigment and the stick's waxy surface, while the background, including the artist's unseen hand and other supplies, completely dissolves into a soft, painterly bokeh of warm, blurred tones, creating a powerful, abstract atmosphere. The image reflects intense, focused creative energy.", + "seed": 8959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_2273_seed_8959_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.42703089136630296, + "vlm_logprob": -0.42703089136630296, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.42703089136630296, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02584064193069935 + ], + "opposite_logprob_tokens": [ + -3.9008407592773438 + ], + "logprob_reward": -0.02584064193069935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02584064193069935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384679526090622, + 0.13277801871299744, + 0.07668489962816238, + 0.18371033668518066, + 0.10718010365962982, + 0.034236788749694824, + 0.16418935358524323, + 0.038803741335868835, + 0.009950481355190277, + 0.11399834603071213 + ], + "credit_map": [ + 0.011281924322247505, + 0.015101755037903786, + 0.023179279640316963, + 0.017011165618896484, + 0.018996017053723335, + 0.015839708968997, + 0.016920730471611023, + 0.020137358456850052, + 0.006839432287961245, + 0.014379742555320263, + 0.007480960804969072, + 0.008170776069164276, + 0.04392119497060776, + 0.011630918830633163, + 0.006462056655436754, + 0.033892933279275894, + 0.006494411267340183, + 0.014215330593287945, + 0.00919748842716217, + 0.009017948992550373, + 0.014234636910259724, + 0.007501974701881409, + 0.00682491110637784, + 0.009198199957609177, + 0.013515492901206017, + 0.06913641095161438, + 0.0067281574010849, + 0.010334971360862255, + 0.024590088054537773, + 0.009469409473240376, + 0.007408114615827799, + 0.042527686804533005, + 0.011446861550211906, + 0.012370824813842773, + 0.010697891935706139, + 0.015003151260316372, + 0.010923449881374836, + 0.0073953415267169476, + 0.020723044872283936, + 0.018619542941451073, + 0.003427511779591441, + 0.0036874141078442335, + 0.005078960210084915, + 0.005073306616395712, + 0.0030380424577742815, + 0.003032457549124956, + 0.005867303349077702, + 0.005031793378293514, + 0.012282194569706917, + 0.050366852432489395, + 0.014081559143960476, + 0.009806284680962563, + 0.024694981053471565, + 0.008392552845180035, + 0.016862936317920685, + 0.02770198881626129, + 0.004671571776270866, + 0.003449589479714632, + 0.0031232438050210476, + 0.003314923495054245, + 0.006569200661033392, + 0.004995559807866812, + 0.005031543783843517, + 0.007648109924048185, + 0.0011657830327749252, + 0.0012492249952629209, + 0.0011911545880138874, + 0.0009434126550331712, + 0.0015868006739765406, + 0.0007936547044664621, + 0.0009380793198943138, + 0.0020823716185986996, + 0.015487397089600563, + 0.015780901536345482, + 0.013218359090387821, + 0.010731359012424946, + 0.011218022555112839, + 0.019890140742063522, + 0.013024168089032173, + 0.014648000709712505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.02584064193069935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029851892963051796 + ], + "opposite_logprob_tokens": [ + -3.5298519134521484 + ], + "logprob_reward": -0.029851892963051796, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029851892963051796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12659861147403717, + 0.14038218557834625, + 0.13097767531871796, + 0.17472045123577118, + 0.06872691214084625, + 0.036595188081264496, + 0.19196736812591553, + 0.02065916918218136, + 0.029790224507451057, + 0.07958221435546875 + ], + "credit_map": [ + 0.009591483511030674, + 0.017662687227129936, + 0.01809309422969818, + 0.013717974536120892, + 0.020147060975432396, + 0.01660270057618618, + 0.015557162463665009, + 0.015226446092128754, + 0.009020856581628323, + 0.014413927681744099, + 0.010160991922020912, + 0.008524506352841854, + 0.042955197393894196, + 0.01330727431923151, + 0.010110949166119099, + 0.03188848868012428, + 0.011413230560719967, + 0.01928066462278366, + 0.013666811399161816, + 0.013450379483401775, + 0.025442376732826233, + 0.010999358259141445, + 0.011458838358521461, + 0.025266028940677643, + 0.012519793584942818, + 0.05810574069619179, + 0.009157818742096424, + 0.015548814088106155, + 0.02322036400437355, + 0.008454253897070885, + 0.010391504503786564, + 0.037322159856557846, + 0.006859468761831522, + 0.009450512938201427, + 0.008322375826537609, + 0.009250814095139503, + 0.007072831969708204, + 0.004769866820424795, + 0.012228763662278652, + 0.010772275738418102, + 0.0033061476424336433, + 0.004238199442625046, + 0.00532055739313364, + 0.00578114902600646, + 0.004403631202876568, + 0.003317939816042781, + 0.005678771995007992, + 0.004548792261630297, + 0.016173211857676506, + 0.05458155274391174, + 0.01885863021016121, + 0.011533133685588837, + 0.028375284746289253, + 0.010176528245210648, + 0.021384580060839653, + 0.030884455889463425, + 0.0021310735028237104, + 0.0022236125078052282, + 0.0021518971771001816, + 0.002358323661610484, + 0.0052613853476941586, + 0.0021146279759705067, + 0.002028163755312562, + 0.0023900847882032394, + 0.003670421661809087, + 0.0030026384629309177, + 0.004009096883237362, + 0.004004979971796274, + 0.003849742701277137, + 0.0029927848372608423, + 0.0028997138142585754, + 0.005360845010727644, + 0.010246867313981056, + 0.010549083352088928, + 0.009767855517566204, + 0.007177602965384722, + 0.008252549916505814, + 0.015768297016620636, + 0.008477907627820969, + 0.00934204924851656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.029851892963051796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103472232818604 + ], + "opposite_logprob_tokens": [ + -0.16034728288650513 + ], + "logprob_reward": -1.9103472232818604, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103472232818604, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07917358726263046, + 0.10017790645360947, + 0.06648079305887222, + 0.2939964532852173, + 0.065053291618824, + 0.020751280710101128, + 0.29028910398483276, + 0.018376588821411133, + 0.030734721571207047, + 0.03496630862355232 + ], + "credit_map": [ + 0.006435789633542299, + 0.011220303364098072, + 0.010362325236201286, + 0.008174475282430649, + 0.010996794328093529, + 0.00793330930173397, + 0.0067780502140522, + 0.017272543162107468, + 0.005009863991290331, + 0.011046184226870537, + 0.005937069188803434, + 0.004777931608259678, + 0.03478426858782768, + 0.010473189875483513, + 0.00626791175454855, + 0.021881485357880592, + 0.006415447220206261, + 0.012072510085999966, + 0.00945117324590683, + 0.008249850943684578, + 0.01063182856887579, + 0.005727088078856468, + 0.007150751072913408, + 0.006782145239412785, + 0.02278236672282219, + 0.09697907418012619, + 0.01557211484760046, + 0.022303447127342224, + 0.0358736626803875, + 0.016026582568883896, + 0.018109707161784172, + 0.06634950637817383, + 0.00775980856269598, + 0.00959628913551569, + 0.005191563628613949, + 0.010955823585391045, + 0.0071230605244636536, + 0.00443532457575202, + 0.0090960543602705, + 0.010895369574427605, + 0.0017372483853250742, + 0.005211332347244024, + 0.0012583488132804632, + 0.0012543406337499619, + 0.0018023953307420015, + 0.0022205926943570375, + 0.0017094491049647331, + 0.005557573400437832, + 0.023965951055288315, + 0.08846576511859894, + 0.026196612045168877, + 0.018463492393493652, + 0.04290458932518959, + 0.01950852759182453, + 0.02959088608622551, + 0.04119327291846275, + 0.002293502679094672, + 0.0018185821827501059, + 0.0015269283903762698, + 0.001721251173876226, + 0.003931892104446888, + 0.0014376362087205052, + 0.0015909149078652263, + 0.004055881407111883, + 0.0031848717480897903, + 0.003080380614846945, + 0.004140641074627638, + 0.004618875216692686, + 0.004561019130051136, + 0.004511659033596516, + 0.003396633779630065, + 0.0032406423706561327, + 0.0034334820229560137, + 0.004336800426244736, + 0.00488531868904829, + 0.005170690361410379, + 0.004145571030676365, + 0.004087904468178749, + 0.004967516288161278, + 0.003939021844416857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -1.9103472232818604 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604650765657425 + ], + "opposite_logprob_tokens": [ + -1.910465121269226 + ], + "logprob_reward": -0.1604650765657425, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604650765657425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07764824479818344, + 0.10235504060983658, + 0.1119142398238182, + 0.3329210877418518, + 0.0427473820745945, + 0.07524183392524719, + 0.16743378341197968, + 0.03224116191267967, + 0.02377082034945488, + 0.03372640907764435 + ], + "credit_map": [ + 0.006784697528928518, + 0.010141023434698582, + 0.011325744912028313, + 0.008221328258514404, + 0.010399868711829185, + 0.00877767987549305, + 0.007591203320771456, + 0.014406691305339336, + 0.004906365182250738, + 0.01067843846976757, + 0.007106125354766846, + 0.0062519703060388565, + 0.034958552569150925, + 0.010964112356305122, + 0.006676803342998028, + 0.02081267163157463, + 0.009264251217246056, + 0.02338036149740219, + 0.015272578224539757, + 0.014697211794555187, + 0.018810395151376724, + 0.007193508092314005, + 0.011362306773662567, + 0.011933630332350731, + 0.02558864839375019, + 0.12247523665428162, + 0.012746955268085003, + 0.019765468314290047, + 0.04357529804110527, + 0.01707977056503296, + 0.014141015708446503, + 0.07754871994256973, + 0.004850068595260382, + 0.007415064610540867, + 0.003994154278188944, + 0.006166638340801001, + 0.004427553154528141, + 0.002683977596461773, + 0.006698845885694027, + 0.006511080078780651, + 0.007030571810901165, + 0.021867243573069572, + 0.004210339859127998, + 0.005083388648927212, + 0.00552061852067709, + 0.006359565071761608, + 0.0039503732696175575, + 0.021219734102487564, + 0.01313274260610342, + 0.050270069390535355, + 0.01309291273355484, + 0.011607853695750237, + 0.024513667449355125, + 0.01102854497730732, + 0.01508287526667118, + 0.02870512567460537, + 0.0036086789332330227, + 0.002825434086844325, + 0.0021961138118058443, + 0.0031575472094118595, + 0.006488613318651915, + 0.0045014116913080215, + 0.0044877477921545506, + 0.004975613672286272, + 0.0020355703309178352, + 0.002502880757674575, + 0.0037798958364874125, + 0.002756709000095725, + 0.0033795086201280355, + 0.004343391861766577, + 0.002188154961913824, + 0.002784709446132183, + 0.003975138999521732, + 0.003826057305559516, + 0.00530998595058918, + 0.004232429899275303, + 0.003479566890746355, + 0.0058265309780836105, + 0.003395490115508437, + 0.0036812061443924904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.1604650765657425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649622090160847 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649622090160847, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649622090160847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291392743587494, + 0.11321719735860825, + 0.1327398419380188, + 0.19838617742061615, + 0.060855843126773834, + 0.02141653560101986, + 0.22684265673160553, + 0.04375196620821953, + 0.02671072632074356, + 0.04693979397416115 + ], + "credit_map": [ + 0.013291646726429462, + 0.014368712902069092, + 0.018065573647618294, + 0.014001937583088875, + 0.015464668162167072, + 0.012385829351842403, + 0.014008757658302784, + 0.027552150189876556, + 0.007994645275175571, + 0.012166300788521767, + 0.0061528850346803665, + 0.006267267279326916, + 0.037213344126939774, + 0.013412175700068474, + 0.006658979691565037, + 0.02335159294307232, + 0.013884222134947777, + 0.025178665295243263, + 0.015004107728600502, + 0.016968930140137672, + 0.017180010676383972, + 0.012078273110091686, + 0.01328303012996912, + 0.019162600859999657, + 0.015015696175396442, + 0.06563897430896759, + 0.010455216281116009, + 0.015024850144982338, + 0.024060972034931183, + 0.01099590864032507, + 0.012285559438169003, + 0.04490901902318001, + 0.007184217683970928, + 0.007548031397163868, + 0.005309032741934061, + 0.012347935698926449, + 0.005680350121110678, + 0.0046606590040028095, + 0.00767036247998476, + 0.010455251671373844, + 0.002443107310682535, + 0.003352964296936989, + 0.0019169935258105397, + 0.0017658721189945936, + 0.002130119828507304, + 0.003268099157139659, + 0.0032762056216597557, + 0.003263173159211874, + 0.019297610968351364, + 0.06748684495687485, + 0.02194613590836525, + 0.014947827905416489, + 0.03343954682350159, + 0.0156299639493227, + 0.02477223053574562, + 0.02932249754667282, + 0.004982580430805683, + 0.004466436803340912, + 0.0036007240414619446, + 0.003173704957589507, + 0.010518432594835758, + 0.0042144362814724445, + 0.0040316530503332615, + 0.008763995952904224, + 0.0024856540840119123, + 0.00275326706469059, + 0.00302195455878973, + 0.004113337956368923, + 0.004117853008210659, + 0.004236470442265272, + 0.002902289852499962, + 0.0030798993539065123, + 0.0062618856318295, + 0.0061973887495696545, + 0.004817471839487553, + 0.004994858987629414, + 0.004979799035936594, + 0.008671936579048634, + 0.0054217553697526455, + 0.005594696383923292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.008649622090160847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, utilizing an incredibly shallow depth of field, focuses with razor-sharp precision on the vibrant, cadmium-red tip of an oil stick. The stick's waxy, creamy texture is visible in minute details, slightly deformed and softened by the application of pressure. As it is vigorously spun and dragged across the rough white canvas, the camera captures the dynamic action in microscopic slow motion. The stick leaves a thick, viscous trail across the canvas while, with controlled violence, flinging countless tiny particles of pure, saturated pigment into the air. These particles form a chaotic yet beautiful spiral, suspended briefly like a colorful galaxy. Bright, warm studio light, emanating from the side, shimmers on the smooth, damp pigment and the stick's waxy surface, while the background, including the artist's unseen hand and other supplies, completely dissolves into a soft, painterly bokeh of warm, blurred tones, creating a powerful, abstract atmosphere. The image reflects intense, focused creative energy.", + "seed": 8960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_11/index_2273_seed_8960_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3057375330477953, + "vlm_logprob": -0.3057375330477953, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.3057375330477953, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014441963285207748 + ], + "opposite_logprob_tokens": [ + -4.639441967010498 + ], + "logprob_reward": -0.014441963285207748, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014441963285207748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05936824902892113, + 0.043552011251449585, + 0.051721617579460144, + 0.07508686184883118, + 0.15318071842193604, + 0.0670180395245552, + 0.09106588363647461, + 0.052146583795547485, + 0.0915667787194252, + 0.31529322266578674 + ], + "credit_map": [ + 0.006826736498624086, + 0.00848854798823595, + 0.00748848170042038, + 0.00676138699054718, + 0.008304977789521217, + 0.010077916085720062, + 0.006339592859148979, + 0.005080610513687134, + 0.009485206566751003, + 0.012228628620505333, + 0.0035422032233327627, + 0.004041614942252636, + 0.003801753744482994, + 0.004047382157295942, + 0.0029701953753829002, + 0.0034350273199379444, + 0.004928927402943373, + 0.006514583248645067, + 0.006355129182338715, + 0.007249331567436457, + 0.0063801975920796394, + 0.00554758682847023, + 0.00858000386506319, + 0.006165856495499611, + 0.01570926047861576, + 0.021468376740813255, + 0.005473891738802195, + 0.004838093649595976, + 0.009210404939949512, + 0.00754989730194211, + 0.004620774183422327, + 0.006216157227754593, + 0.015782516449689865, + 0.022826828062534332, + 0.006762254051864147, + 0.03365028277039528, + 0.0367693193256855, + 0.016304589807987213, + 0.009421421214938164, + 0.011663496494293213, + 0.007229501847177744, + 0.011353782378137112, + 0.006119585130363703, + 0.004031885880976915, + 0.012446516193449497, + 0.008073321543633938, + 0.009977372363209724, + 0.0077860718593001366, + 0.011478028260171413, + 0.028925010934472084, + 0.007450278848409653, + 0.01881122961640358, + 0.007663739379495382, + 0.006188035011291504, + 0.005360701121389866, + 0.005188866518437862, + 0.007080088369548321, + 0.014332695864140987, + 0.0037426440976560116, + 0.00488064531236887, + 0.006666481029242277, + 0.00700676254928112, + 0.004897053353488445, + 0.0035402157809585333, + 0.01136963814496994, + 0.006825128570199013, + 0.006067374721169472, + 0.02818410098552704, + 0.008648407645523548, + 0.009997096844017506, + 0.011024470441043377, + 0.009450563229620457, + 0.01962016336619854, + 0.033236052840948105, + 0.02313612587749958, + 0.03760815039277077, + 0.07175043225288391, + 0.04987910017371178, + 0.04819190129637718, + 0.031871285289525986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.014441963285207748 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223858714103699 + ], + "opposite_logprob_tokens": [ + -4.264224052429199 + ], + "logprob_reward": -0.014223858714103699, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223858714103699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0672125294804573, + 0.13812261819839478, + 0.06696151196956635, + 0.09563877433538437, + 0.1589776873588562, + 0.08385859429836273, + 0.1296653300523758, + 0.1761162132024765, + 0.046174343675374985, + 0.037272363901138306 + ], + "credit_map": [ + 0.007082311902195215, + 0.009750209748744965, + 0.007654710207134485, + 0.007755163591355085, + 0.011176856234669685, + 0.012638023123145103, + 0.0055110459215939045, + 0.005644209682941437, + 0.03711607679724693, + 0.04335794225335121, + 0.008665243163704872, + 0.011240687221288681, + 0.012940164655447006, + 0.009212521836161613, + 0.007279396057128906, + 0.008310595527291298, + 0.005918548442423344, + 0.009736409410834312, + 0.006651422940194607, + 0.010527239181101322, + 0.008251325227320194, + 0.010609367862343788, + 0.009629085659980774, + 0.005638122092932463, + 0.011806530877947807, + 0.014300390146672726, + 0.00964410975575447, + 0.01530136726796627, + 0.01253390684723854, + 0.011677984148263931, + 0.009718073531985283, + 0.010656419210135937, + 0.009113520383834839, + 0.04238997399806976, + 0.007944402284920216, + 0.011837431229650974, + 0.05441909283399582, + 0.016021303832530975, + 0.008861042559146881, + 0.008390918374061584, + 0.01734812930226326, + 0.024365603923797607, + 0.007213601376861334, + 0.006906304974108934, + 0.00563282985240221, + 0.009670177474617958, + 0.005760644096881151, + 0.00696130795404315, + 0.0247186291962862, + 0.032288938760757446, + 0.008507807739078999, + 0.017544077709317207, + 0.015517138876020908, + 0.013619674369692802, + 0.009550229646265507, + 0.007918830029666424, + 0.036519281566143036, + 0.06843411177396774, + 0.010312361642718315, + 0.018690919503569603, + 0.012840799055993557, + 0.01380824577063322, + 0.00830316822975874, + 0.007207334972918034, + 0.0051549929194152355, + 0.004260747227817774, + 0.002820107853040099, + 0.01017450075596571, + 0.006938902661204338, + 0.006781309377402067, + 0.005051913671195507, + 0.004991871304810047, + 0.0018976939609274268, + 0.0038948352448642254, + 0.0032969641033560038, + 0.004251275211572647, + 0.0085643595084548, + 0.005445521324872971, + 0.006085474509745836, + 0.0038362417835742235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.014223858714103699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063094854354858 + ], + "opposite_logprob_tokens": [ + -0.28130948543548584 + ], + "logprob_reward": -1.4063094854354858, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063094854354858, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04776792973279953, + 0.08893300592899323, + 0.05635660141706467, + 0.1132005900144577, + 0.05876057967543602, + 0.019643180072307587, + 0.47051265835762024, + 0.09238361567258835, + 0.019283276051282883, + 0.03315853700041771 + ], + "credit_map": [ + 0.006163300946354866, + 0.0059618232771754265, + 0.006999588571488857, + 0.0044629937037825584, + 0.005335424095392227, + 0.009026921354234219, + 0.00516610499471426, + 0.004651772323995829, + 0.02169700339436531, + 0.02551218867301941, + 0.0059995208866894245, + 0.005489194765686989, + 0.01052685547620058, + 0.0069571807980537415, + 0.008147187530994415, + 0.004603874869644642, + 0.004783466458320618, + 0.010677170008420944, + 0.004963114857673645, + 0.008164217695593834, + 0.007192761637270451, + 0.00818691961467266, + 0.006150885485112667, + 0.006238067988306284, + 0.01577587053179741, + 0.020724892616271973, + 0.011665674857795238, + 0.010488876141607761, + 0.01463436521589756, + 0.012113983742892742, + 0.014058523811399937, + 0.013738404028117657, + 0.005664254073053598, + 0.006843044422566891, + 0.0036767933052033186, + 0.014003943651914597, + 0.01441191416233778, + 0.006477117072790861, + 0.003702839370816946, + 0.003980671986937523, + 0.0033771051093935966, + 0.004961400292813778, + 0.002067794091999531, + 0.0020192116498947144, + 0.0018577048322185874, + 0.0020535059738904238, + 0.0018311867024749517, + 0.0014752700226381421, + 0.03365639969706535, + 0.14186452329158783, + 0.024274546653032303, + 0.16373209655284882, + 0.03425728157162666, + 0.022320102900266647, + 0.031494010239839554, + 0.01891370676457882, + 0.007135517429560423, + 0.01127545814961195, + 0.01114609558135271, + 0.014471792615950108, + 0.015034771524369717, + 0.013902470469474792, + 0.013438837602734566, + 0.005978677421808243, + 0.002046892186626792, + 0.0018558157607913017, + 0.001234302413649857, + 0.004677486605942249, + 0.0015688342973589897, + 0.0027757808566093445, + 0.0020663090981543064, + 0.003057853551581502, + 0.00320122460834682, + 0.004691583104431629, + 0.0028182563837617636, + 0.003889931831508875, + 0.0068290443159639835, + 0.004452428314834833, + 0.0038569641765207052, + 0.003419103566557169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -1.4063094854354858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907207101583481 + ], + "opposite_logprob_tokens": [ + -2.4640719890594482 + ], + "logprob_reward": -0.08907207101583481, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907207101583481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04440182074904442, + 0.07228419929742813, + 0.08827227354049683, + 0.08975549787282944, + 0.09302559494972229, + 0.03745635971426964, + 0.3960596024990082, + 0.1325061172246933, + 0.02920534834265709, + 0.017033182084560394 + ], + "credit_map": [ + 0.004904173314571381, + 0.005985284689813852, + 0.005880902986973524, + 0.004687957465648651, + 0.004537093453109264, + 0.009929182007908821, + 0.004591251257807016, + 0.0038859755732119083, + 0.0081837372854352, + 0.009273368865251541, + 0.009686319157481194, + 0.00902602355927229, + 0.01144133135676384, + 0.008967799134552479, + 0.009772375226020813, + 0.005933242850005627, + 0.007913745939731598, + 0.01401471346616745, + 0.00842096284031868, + 0.01211764570325613, + 0.012431773357093334, + 0.013992038555443287, + 0.010087830014526844, + 0.009293563663959503, + 0.009923088364303112, + 0.012365867383778095, + 0.009979915805161, + 0.010274535045027733, + 0.011916319839656353, + 0.010534921661019325, + 0.01249739807099104, + 0.012263446114957333, + 0.0046356674283742905, + 0.01989019848406315, + 0.003861483884975314, + 0.013360034674406052, + 0.03430105373263359, + 0.007437181659042835, + 0.0043525248765945435, + 0.00518744857981801, + 0.006970949936658144, + 0.009618769399821758, + 0.002866766881197691, + 0.0033947802148759365, + 0.002985310275107622, + 0.0046768770553171635, + 0.0028601777739822865, + 0.004082728177309036, + 0.024706441909074783, + 0.11526575684547424, + 0.021248476579785347, + 0.14546212553977966, + 0.02777729742228985, + 0.02016831375658512, + 0.028456058353185654, + 0.012975125573575497, + 0.02706781215965748, + 0.0485386997461319, + 0.007629921659827232, + 0.010480262339115143, + 0.010439039207994938, + 0.011727060191333294, + 0.010012880899012089, + 0.006610441487282515, + 0.003033372573554516, + 0.0030104462057352066, + 0.0022511000279337168, + 0.0033179819583892822, + 0.0031070513650774956, + 0.004550461657345295, + 0.003396657994017005, + 0.006538276560604572, + 0.0017081338446587324, + 0.002549226861447096, + 0.001463885186240077, + 0.0017428472638130188, + 0.0035862047225236893, + 0.0017979281255975366, + 0.002519409405067563, + 0.0016655467916280031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.08907207101583481 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of an oil stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the oil stick tip cadmium-red with a visible waxy, creamy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the oil stick being vigorously spun and dragged across a rough white canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny particles of saturated pigment fling into the air forming a chaotic spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dissolved into a soft, painterly bokeh of warm, blurred tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640286788344383 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640286788344383, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640286788344383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03919686749577522, + 0.09642230719327927, + 0.04796668887138367, + 0.041903335601091385, + 0.1544749140739441, + 0.06753814220428467, + 0.31615397334098816, + 0.10069190710783005, + 0.06711750477552414, + 0.06853438168764114 + ], + "credit_map": [ + 0.00534076476469636, + 0.004116090014576912, + 0.005004857201129198, + 0.0037002924364060163, + 0.005480695050209761, + 0.007854718714952469, + 0.0035296606365591288, + 0.004169787745922804, + 0.022029254585504532, + 0.02660263516008854, + 0.007443523965775967, + 0.006984316278249025, + 0.01131216436624527, + 0.007951444946229458, + 0.009125621058046818, + 0.0049733491614460945, + 0.004440292250365019, + 0.008612215518951416, + 0.003979260101914406, + 0.007219715043902397, + 0.006089147180318832, + 0.006598754320293665, + 0.005211168434470892, + 0.005816134624183178, + 0.00570907024666667, + 0.007223892025649548, + 0.004349129740148783, + 0.003868749598041177, + 0.005627300590276718, + 0.004152393434196711, + 0.005636490881443024, + 0.005336306989192963, + 0.026120895519852638, + 0.014714266173541546, + 0.013126129284501076, + 0.030473949387669563, + 0.017528947442770004, + 0.022947469726204872, + 0.01477630902081728, + 0.014786952175199986, + 0.01388689037412405, + 0.0239958968013525, + 0.004785715136677027, + 0.004704824183136225, + 0.0053450255654752254, + 0.006475777365267277, + 0.004336964339017868, + 0.004007047973573208, + 0.034780047833919525, + 0.1119919940829277, + 0.014707004651427269, + 0.09889473021030426, + 0.014352413825690746, + 0.013587829656898975, + 0.01664191484451294, + 0.011198029853403568, + 0.014394092373549938, + 0.022035742178559303, + 0.010951611213386059, + 0.01257608737796545, + 0.013322949409484863, + 0.011528895236551762, + 0.009516209363937378, + 0.006366325542330742, + 0.005782521795481443, + 0.006535950116813183, + 0.005096933804452419, + 0.015361038967967033, + 0.00934044923633337, + 0.008956068195402622, + 0.007939232513308525, + 0.008105309680104256, + 0.005477671045809984, + 0.008236048743128777, + 0.005410814192146063, + 0.007998916320502758, + 0.014708454720675945, + 0.01036018319427967, + 0.00980277732014656, + 0.006539515219628811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.004640286788344383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_tensor([4179, 4179])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_tensor([4179, 4179])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c1c8680c611f1118a9932e8d8921b75362da6ef2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_tensor([4179, 4179])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static cinematic shot with thick film grain depicts a massive, dark-bodied Galapagos centipede, its form coiled in immense tension as it prepares to impact a mossy rock. The entire scene is rendered in a distinctive GTA art style, defined by bold black outlines and a saturated pop-art palette. The centipede's segmented exoskeleton has a lustrous, deep charcoal sheen that catches the harsh, low-angle light, creating vibrant, stylized highlights. Its countless, sharp legs are a startling electric yellow, contrasting starkly against its black body. Its head rises menacingly, its long tentacles twitching erratically, while its powerful form flares slightly, gleaming with malevolent promise. The creature rests on a large, weathered rock, its texture simplified to a graphic plane and covered in a supersaturated, almost jagged patch of moss. In the background, a stylized jungle environment dissolves into soft focus beneath a dramatic, synthetic sunset sky, painted in hues of deep purple and fiery orange. Intense directional lighting casts sharp-edged graphic shadows that accentuate the creature’s awe-inspiring silhouette and give this menacing creature an unexpectedly vibrant and stylized digital appearance, creating a uniquely tense and visually appealing atmosphere.", + "seed": 12609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_4179_seed_12609_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010901203332468867, + "vlm_logprob": -0.010901203332468867, + "vlm_yes_no_margin": 4.8249999523162845, + "vlm_token_credit": -0.010901203332468867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019527103751897812 + ], + "opposite_logprob_tokens": [ + -4.019526958465576 + ], + "logprob_reward": -0.019527103751897812, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.019527103751897812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21856118738651276, + 0.05725511536002159, + 0.04229963198304176, + 0.03130241483449936, + 0.04762065038084984, + 0.04004230722784996, + 0.34414756298065186, + 0.056314777582883835, + 0.04786744713783264, + 0.11458887904882431 + ], + "credit_map": [ + 0.020311465486884117, + 0.019833048805594444, + 0.02840554714202881, + 0.03649330511689186, + 0.017148923128843307, + 0.035663437098264694, + 0.031859781593084335, + 0.02884570322930813, + 0.0057738171890378, + 0.005951608531177044, + 0.00858668889850378, + 0.009136371314525604, + 0.006036596838384867, + 0.007655596360564232, + 0.0074318088591098785, + 0.006682633887976408, + 0.005401051603257656, + 0.003504474414512515, + 0.004157889634370804, + 0.005310129839926958, + 0.004385370761156082, + 0.006278662476688623, + 0.006597865838557482, + 0.006664188113063574, + 0.004708948545157909, + 0.0035239001736044884, + 0.0037316882517188787, + 0.0055328551679849625, + 0.0036543451715260744, + 0.002718528965488076, + 0.0030924712773412466, + 0.004339682869613171, + 0.0057968092150986195, + 0.004156478215008974, + 0.005154292099177837, + 0.006066866684705019, + 0.00392782362177968, + 0.006885649170726538, + 0.005460508167743683, + 0.010172227397561073, + 0.005110728554427624, + 0.004417293705046177, + 0.004567039664834738, + 0.006993788760155439, + 0.004762674681842327, + 0.0034775049425661564, + 0.005585840903222561, + 0.005127439275383949, + 0.03865096718072891, + 0.03718499466776848, + 0.04781310260295868, + 0.0310678631067276, + 0.025172121822834015, + 0.09231817722320557, + 0.036219481378793716, + 0.035720858722925186, + 0.008655824698507786, + 0.005372250452637672, + 0.007087586913257837, + 0.0062926518730819225, + 0.0044004409573972225, + 0.009419572539627552, + 0.006137165706604719, + 0.008949287235736847, + 0.004907740745693445, + 0.0030078114941716194, + 0.008508866652846336, + 0.004754857625812292, + 0.0045919232070446014, + 0.01268305815756321, + 0.003775532590225339, + 0.005637660156935453, + 0.014115460216999054, + 0.017718205228447914, + 0.014879868365824223, + 0.013750847429037094, + 0.013631447218358517, + 0.011869478970766068, + 0.013504798524081707, + 0.015118778683245182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.019527103751897812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602276548743248 + ], + "opposite_logprob_tokens": [ + -3.895602226257324 + ], + "logprob_reward": -0.020602276548743248, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602276548743248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11842343211174011, + 0.28999289870262146, + 0.11941485852003098, + 0.07562151551246643, + 0.026361286640167236, + 0.0320214107632637, + 0.1539277583360672, + 0.08699125051498413, + 0.05176969990134239, + 0.04547588527202606 + ], + "credit_map": [ + 0.01651253178715706, + 0.012263978831470013, + 0.01666313409805298, + 0.016063006594777107, + 0.013245776295661926, + 0.015586966648697853, + 0.01477240864187479, + 0.013315618969500065, + 0.02485431358218193, + 0.028149640187621117, + 0.04045914113521576, + 0.041675180196762085, + 0.026648636907339096, + 0.03601379692554474, + 0.054537225514650345, + 0.037654951214790344, + 0.013618262484669685, + 0.01200827956199646, + 0.013016513548791409, + 0.015619450248777866, + 0.020171208307147026, + 0.012548833154141903, + 0.016562344506382942, + 0.01586996018886566, + 0.0072994958609342575, + 0.0068078357726335526, + 0.009695370681583881, + 0.01812811940908432, + 0.0080918213352561, + 0.007142479531466961, + 0.008510624058544636, + 0.009945763275027275, + 0.003051553387194872, + 0.002647900255396962, + 0.0025857139844447374, + 0.00388832064345479, + 0.0025313030928373337, + 0.0039426954463124275, + 0.0034031469840556383, + 0.004310651682317257, + 0.003657593857496977, + 0.0035301847383379936, + 0.004988443106412888, + 0.006006327923387289, + 0.00274963048286736, + 0.0025201099924743176, + 0.004662625957280397, + 0.0039064944721758366, + 0.017520910128951073, + 0.010521797463297844, + 0.03428918123245239, + 0.01481492817401886, + 0.009785610251128674, + 0.037353891879320145, + 0.014548240229487419, + 0.01509319432079792, + 0.011664265766739845, + 0.010768545791506767, + 0.009136611595749855, + 0.010472034104168415, + 0.008636974729597569, + 0.015919188037514687, + 0.008372075855731964, + 0.012021547183394432, + 0.005496853496879339, + 0.003102169604972005, + 0.01128655206412077, + 0.004936075769364834, + 0.0052962820045650005, + 0.013608294539153576, + 0.0029758319724351168, + 0.005067639984190464, + 0.009500394575297832, + 0.004800924565643072, + 0.005112840794026852, + 0.004205557517707348, + 0.0051342095248401165, + 0.009805978275835514, + 0.003198644844815135, + 0.003717334009706974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.020602276548743248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24705560505390167, + 0.09013430774211884, + 0.024094106629490852, + 0.0720125064253807, + 0.024355297908186913, + 0.07218904793262482, + 0.15727664530277252, + 0.03668995946645737, + 0.19370031356811523, + 0.08249218761920929 + ], + "credit_map": [ + 0.032137151807546616, + 0.02461869642138481, + 0.030330462381243706, + 0.02164638042449951, + 0.024591945111751556, + 0.041227806359529495, + 0.02856956422328949, + 0.043933600187301636, + 0.009766383096575737, + 0.008971694856882095, + 0.011658611707389355, + 0.00910306628793478, + 0.008754713460803032, + 0.009885743260383606, + 0.01680249720811844, + 0.015191606245934963, + 0.0023775026202201843, + 0.003680658293887973, + 0.0027951246593147516, + 0.0030666859820485115, + 0.0023366534151136875, + 0.002626494737342, + 0.0030982824973762035, + 0.004112705122679472, + 0.0067833526991307735, + 0.006129852030426264, + 0.009008850902318954, + 0.008484394289553165, + 0.010270027443766594, + 0.008625839836895466, + 0.010582688264548779, + 0.012127500027418137, + 0.002682980615645647, + 0.003168658120557666, + 0.0026129158213734627, + 0.0026948268059641123, + 0.0022727095056325197, + 0.0028046395163983107, + 0.003351394319906831, + 0.004767172504216433, + 0.009006036445498466, + 0.010857845656573772, + 0.009667142294347286, + 0.010816081427037716, + 0.006042038090527058, + 0.00624008709564805, + 0.008936086669564247, + 0.010623738169670105, + 0.017454279586672783, + 0.008770452812314034, + 0.03440345823764801, + 0.015656262636184692, + 0.010359333828091621, + 0.039207931607961655, + 0.015686189755797386, + 0.015738731250166893, + 0.005421411246061325, + 0.0037483747582882643, + 0.004316756036132574, + 0.003789345733821392, + 0.003896385431289673, + 0.007371814921498299, + 0.0033882076386362314, + 0.004757661372423172, + 0.02180539071559906, + 0.010427424684166908, + 0.04153132438659668, + 0.01590997539460659, + 0.019066059961915016, + 0.057179998606443405, + 0.010145137086510658, + 0.01763501763343811, + 0.009907987900078297, + 0.012498172000050545, + 0.008241739124059677, + 0.009939992800354958, + 0.01132248155772686, + 0.011381975375115871, + 0.007469402626156807, + 0.011730432510375977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005946207791566849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944074597209692 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944074597209692, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944074597209692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11125025153160095, + 0.2524184286594391, + 0.047835588455200195, + 0.06028304249048233, + 0.04470399394631386, + 0.08777157962322235, + 0.1630571186542511, + 0.11701475083827972, + 0.03288630023598671, + 0.0827789455652237 + ], + "credit_map": [ + 0.013749354518949986, + 0.013593955896794796, + 0.014221719466149807, + 0.014907851815223694, + 0.012790117412805557, + 0.01961725950241089, + 0.010282622650265694, + 0.012087369337677956, + 0.01924464851617813, + 0.0245891734957695, + 0.03351632505655289, + 0.033671531826257706, + 0.0244916882365942, + 0.03545409440994263, + 0.051570307463407516, + 0.029880661517381668, + 0.004962525330483913, + 0.004533525090664625, + 0.00534317921847105, + 0.005652034189552069, + 0.007598592899739742, + 0.005856404080986977, + 0.007195374928414822, + 0.006693952716886997, + 0.006446478422731161, + 0.0065600331872701645, + 0.007614292670041323, + 0.009329227730631828, + 0.005870784632861614, + 0.006062396336346865, + 0.005970681086182594, + 0.012429147027432919, + 0.0053617204539477825, + 0.005020704586058855, + 0.005678832530975342, + 0.005755783524364233, + 0.004432665184140205, + 0.0041036359034478664, + 0.006523690186440945, + 0.007826962508261204, + 0.008541355840861797, + 0.019147424027323723, + 0.008605167269706726, + 0.017541619017720222, + 0.005623252131044865, + 0.006351084914058447, + 0.011621157638728619, + 0.010340522974729538, + 0.01921779476106167, + 0.023913534358143806, + 0.019839538261294365, + 0.017504045739769936, + 0.012434638105332851, + 0.0402560792863369, + 0.017769457772374153, + 0.012122014537453651, + 0.015514841303229332, + 0.018033679574728012, + 0.00920884683728218, + 0.010270972736179829, + 0.01212327554821968, + 0.03014574944972992, + 0.011337827891111374, + 0.010379564948379993, + 0.003841266967356205, + 0.00250837835483253, + 0.00911406334489584, + 0.004218929447233677, + 0.0029148876201361418, + 0.002559896092861891, + 0.003075851360335946, + 0.004653026349842548, + 0.011603464372456074, + 0.010134024545550346, + 0.008686217479407787, + 0.013122807256877422, + 0.006704399827867746, + 0.01720558851957321, + 0.007012129761278629, + 0.008310318924486637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.005944074597209692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486353972926736 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.002486353972926736, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486353972926736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35130712389945984, + 0.08777095377445221, + 0.09222247451543808, + 0.060977671295404434, + 0.05737864598631859, + 0.1947038471698761, + 0.04174832999706268, + 0.03600717708468437, + 0.03372151777148247, + 0.044162310659885406 + ], + "credit_map": [ + 0.047590821981430054, + 0.032246995717287064, + 0.04216741397976875, + 0.03382846340537071, + 0.0398116335272789, + 0.06777401268482208, + 0.033838365226984024, + 0.054049376398324966, + 0.007926139049232006, + 0.009087616577744484, + 0.012490271590650082, + 0.008420130237936974, + 0.011878298595547676, + 0.009293434210121632, + 0.01661239191889763, + 0.012062666937708855, + 0.00795894768089056, + 0.008014680817723274, + 0.010210671462118626, + 0.00987507775425911, + 0.016100553795695305, + 0.007700357120484114, + 0.016350992023944855, + 0.016011187806725502, + 0.009496200829744339, + 0.008905366063117981, + 0.006485329009592533, + 0.005791991483420134, + 0.007744873408228159, + 0.009119383990764618, + 0.005635559093207121, + 0.007798962760716677, + 0.007593881338834763, + 0.007047201972454786, + 0.007173480000346899, + 0.007465752307325602, + 0.006651715375483036, + 0.006190855521708727, + 0.006702616345137358, + 0.008553137071430683, + 0.021608781069517136, + 0.0313958004117012, + 0.031245671212673187, + 0.02720923349261284, + 0.013533011078834534, + 0.023116271942853928, + 0.0195973739027977, + 0.026997683569788933, + 0.004166888538748026, + 0.004341535735875368, + 0.010487535037100315, + 0.005605088546872139, + 0.0027974280528724194, + 0.006400339305400848, + 0.003976135980337858, + 0.003973374143242836, + 0.004771646112203598, + 0.0040429700165987015, + 0.0039861006662249565, + 0.004888467025011778, + 0.005068500991910696, + 0.005483133718371391, + 0.003575634676963091, + 0.004190722480416298, + 0.004758819006383419, + 0.003437980078160763, + 0.004781755618751049, + 0.003474462777376175, + 0.003384208306670189, + 0.004902681335806847, + 0.003090072888880968, + 0.005891535896807909, + 0.004790660459548235, + 0.005113452207297087, + 0.005536304321140051, + 0.006247548386454582, + 0.00800558365881443, + 0.006866712588816881, + 0.0033381013199687004, + 0.004263943526893854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.002486353972926736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static cinematic shot with thick film grain depicts a massive, dark-bodied Galapagos centipede, its form coiled in immense tension as it prepares to impact a mossy rock. The entire scene is rendered in a distinctive GTA art style, defined by bold black outlines and a saturated pop-art palette. The centipede's segmented exoskeleton has a lustrous, deep charcoal sheen that catches the harsh, low-angle light, creating vibrant, stylized highlights. Its countless, sharp legs are a startling electric yellow, contrasting starkly against its black body. Its head rises menacingly, its long tentacles twitching erratically, while its powerful form flares slightly, gleaming with malevolent promise. The creature rests on a large, weathered rock, its texture simplified to a graphic plane and covered in a supersaturated, almost jagged patch of moss. In the background, a stylized jungle environment dissolves into soft focus beneath a dramatic, synthetic sunset sky, painted in hues of deep purple and fiery orange. Intense directional lighting casts sharp-edged graphic shadows that accentuate the creature’s awe-inspiring silhouette and give this menacing creature an unexpectedly vibrant and stylized digital appearance, creating a uniquely tense and visually appealing atmosphere.", + "seed": 12610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_4179_seed_12610_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024623708752915262, + "vlm_logprob": -0.024623708752915262, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.024623708752915262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04945031926035881 + ], + "opposite_logprob_tokens": [ + -3.049450397491455 + ], + "logprob_reward": -0.04945031926035881, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04945031926035881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431528627872467, + 0.04733256250619888, + 0.25372985005378723, + 0.0319584459066391, + 0.0324045866727829, + 0.016893304884433746, + 0.2532314658164978, + 0.043824464082717896, + 0.14277227222919464, + 0.03470010682940483 + ], + "credit_map": [ + 0.023875927552580833, + 0.013005583547055721, + 0.023067358881235123, + 0.01661837473511696, + 0.017271503806114197, + 0.017630258575081825, + 0.019790921360254288, + 0.011892943643033504, + 0.008878975175321102, + 0.005091085564345121, + 0.009293046779930592, + 0.005594126880168915, + 0.00474232854321599, + 0.004565701819956303, + 0.005579700227826834, + 0.003587601473554969, + 0.038521524518728256, + 0.018788829445838928, + 0.03647506609559059, + 0.036175116896629333, + 0.039176665246486664, + 0.03210578113794327, + 0.029606137424707413, + 0.022880738601088524, + 0.005961090791970491, + 0.0050474852323532104, + 0.005483210552483797, + 0.0028094183653593063, + 0.003511351766064763, + 0.004035584628582001, + 0.002646772190928459, + 0.0024635354056954384, + 0.006648652721196413, + 0.0026882016099989414, + 0.006329466123133898, + 0.0035471206065267324, + 0.004582313355058432, + 0.0038693235255777836, + 0.0021193078719079494, + 0.002620202489197254, + 0.002898433245718479, + 0.002109172288328409, + 0.0029358668252825737, + 0.0029708007350564003, + 0.0019414747366681695, + 0.0013364795595407486, + 0.001393567188642919, + 0.0013075111201032996, + 0.08067134767770767, + 0.018516987562179565, + 0.07933557033538818, + 0.019035154953598976, + 0.01730763167142868, + 0.01658831536769867, + 0.011584015563130379, + 0.01019246131181717, + 0.005535876378417015, + 0.0029156787786632776, + 0.0066428473219275475, + 0.005410484503954649, + 0.007795570883899927, + 0.0037749800831079483, + 0.004089851398020983, + 0.007659179158508778, + 0.03513336181640625, + 0.01962059549987316, + 0.030966518446803093, + 0.012137330137193203, + 0.011606229469180107, + 0.014455829747021198, + 0.008421681821346283, + 0.010430742055177689, + 0.004649383947253227, + 0.0034499072935432196, + 0.004639119375497103, + 0.006152600049972534, + 0.0055592963472008705, + 0.003462405176833272, + 0.004327598959207535, + 0.002459799638018012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.04945031926035881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207360327243805 + ], + "opposite_logprob_tokens": [ + -2.8120737075805664 + ], + "logprob_reward": -0.06207360327243805, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207360327243805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14914721250534058, + 0.06898831576108932, + 0.22032247483730316, + 0.07516779005527496, + 0.04432402178645134, + 0.05110074579715729, + 0.13254937529563904, + 0.06730520725250244, + 0.13051116466522217, + 0.060583699494600296 + ], + "credit_map": [ + 0.032793253660202026, + 0.010819484479725361, + 0.035567063838243484, + 0.023745188489556313, + 0.011860296130180359, + 0.011208727024495602, + 0.015909820795059204, + 0.0072433664463460445, + 0.014866233803331852, + 0.008875035680830479, + 0.013510454446077347, + 0.0075264074839651585, + 0.0074653602205216885, + 0.005999715067446232, + 0.005523432977497578, + 0.005221668630838394, + 0.059889473021030426, + 0.015880052000284195, + 0.05974414199590683, + 0.022939741611480713, + 0.015304828062653542, + 0.013134310953319073, + 0.01908079721033573, + 0.014349115081131458, + 0.012544757686555386, + 0.012956398539245129, + 0.01077950932085514, + 0.00546692218631506, + 0.009895369410514832, + 0.010166928172111511, + 0.0075036995112895966, + 0.005854199640452862, + 0.00947204977273941, + 0.004313976503908634, + 0.009622366167604923, + 0.004450272768735886, + 0.004349561408162117, + 0.005533994175493717, + 0.0029312269762158394, + 0.003650570521131158, + 0.011354561895132065, + 0.005389503203332424, + 0.009382392279803753, + 0.006963276769965887, + 0.005273707676678896, + 0.003909063059836626, + 0.004878857173025608, + 0.003949381876736879, + 0.027776889503002167, + 0.012511027045547962, + 0.027811573818325996, + 0.013792476616799831, + 0.015351950190961361, + 0.010946876369416714, + 0.013500242494046688, + 0.010858333669602871, + 0.007970249280333519, + 0.009928305633366108, + 0.008486228063702583, + 0.005793067626655102, + 0.00963237788528204, + 0.00927469041198492, + 0.006390357855707407, + 0.0098299290984869, + 0.034341443330049515, + 0.014949760399758816, + 0.030981851741671562, + 0.014793981797993183, + 0.013437465764582157, + 0.00787847489118576, + 0.007540447637438774, + 0.006587729789316654, + 0.010477405972778797, + 0.005735437385737896, + 0.008846992626786232, + 0.009785952977836132, + 0.00826072134077549, + 0.0062405867502093315, + 0.007197829429060221, + 0.004038762766867876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.06207360327243805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046355403028428555 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.0046355403028428555, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046355403028428555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13769316673278809, + 0.03156181797385216, + 0.09644661098718643, + 0.14262162148952484, + 0.017339831218123436, + 0.03427152335643768, + 0.28584232926368713, + 0.03659898042678833, + 0.19146466255187988, + 0.02615951932966709 + ], + "credit_map": [ + 0.024634428322315216, + 0.015157748945057392, + 0.025387555360794067, + 0.022049812600016594, + 0.011214915663003922, + 0.014770425856113434, + 0.015244320966303349, + 0.009233961813151836, + 0.006907429546117783, + 0.003675836604088545, + 0.005919128190726042, + 0.002515080850571394, + 0.0027793999761343002, + 0.003970328718423843, + 0.0032992458436638117, + 0.0024953652173280716, + 0.020518528297543526, + 0.010601164773106575, + 0.018123837187886238, + 0.01053281594067812, + 0.010219631716609001, + 0.00922843161970377, + 0.009380477480590343, + 0.007841724902391434, + 0.03336922079324722, + 0.019289197400212288, + 0.027945537120103836, + 0.01089218258857727, + 0.013890753500163555, + 0.014739449135959148, + 0.009585816413164139, + 0.012909465469419956, + 0.0037004484329372644, + 0.0020349021069705486, + 0.0036669564433395863, + 0.0016906108940020204, + 0.002331660594791174, + 0.0016907969256862998, + 0.0011729769175872207, + 0.001051478786394, + 0.006007117219269276, + 0.004523001611232758, + 0.004694770090281963, + 0.005089533980935812, + 0.003401880618184805, + 0.00411086343228817, + 0.0033695383463054895, + 0.0030748173594474792, + 0.09946570545434952, + 0.02549973502755165, + 0.09709127992391586, + 0.019132554531097412, + 0.010798418894410133, + 0.013701584190130234, + 0.01134251244366169, + 0.008810543455183506, + 0.004099695943295956, + 0.0032432388979941607, + 0.004801977891474962, + 0.002987784566357732, + 0.004692846909165382, + 0.0040788971818983555, + 0.0038530754391103983, + 0.008841465227305889, + 0.04664141312241554, + 0.02610178105533123, + 0.04064005985856056, + 0.01673562452197075, + 0.015906283631920815, + 0.01958509534597397, + 0.011134096421301365, + 0.014720299281179905, + 0.004292386118322611, + 0.004342433996498585, + 0.0038457342889159918, + 0.003920983523130417, + 0.0026723321061581373, + 0.0023891576565802097, + 0.0030018675606697798, + 0.0016946264076977968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.0046355403028428555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052477470599114895 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052477470599114895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052477470599114895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13692361116409302, + 0.13422109186649323, + 0.12188003957271576, + 0.1358533501625061, + 0.05673057958483696, + 0.07541245967149734, + 0.11134664714336395, + 0.05553394556045532, + 0.09457235038280487, + 0.07752592861652374 + ], + "credit_map": [ + 0.02938390150666237, + 0.015056652016937733, + 0.027456367388367653, + 0.02014359086751938, + 0.009987265802919865, + 0.00922902300953865, + 0.011772158555686474, + 0.013894644565880299, + 0.02613435871899128, + 0.011545872315764427, + 0.022452451288700104, + 0.01931094378232956, + 0.011208334937691689, + 0.010972674936056137, + 0.01777208410203457, + 0.014824379235506058, + 0.02120155468583107, + 0.01071880478411913, + 0.022486848756670952, + 0.01625928282737732, + 0.011912178248167038, + 0.015223459340631962, + 0.014142672531306744, + 0.009935230016708374, + 0.0212567038834095, + 0.011319058015942574, + 0.019173165783286095, + 0.013410368002951145, + 0.020984217524528503, + 0.015030747279524803, + 0.016678737476468086, + 0.01800035871565342, + 0.006865144241601229, + 0.009342378936707973, + 0.006947894115000963, + 0.005956804379820824, + 0.008276277221739292, + 0.007821447215974331, + 0.0066133541986346245, + 0.004907282069325447, + 0.018099764361977577, + 0.00833672471344471, + 0.01538314949721098, + 0.01011950708925724, + 0.007253541145473719, + 0.005628050770610571, + 0.006471904460340738, + 0.004119819961488247, + 0.038923829793930054, + 0.008183922618627548, + 0.0381852351129055, + 0.007735195569694042, + 0.0042993430979549885, + 0.004827823489904404, + 0.00513723399490118, + 0.004054060205817223, + 0.007791506126523018, + 0.007669099606573582, + 0.008612935431301594, + 0.005328374449163675, + 0.007205082103610039, + 0.006745036691427231, + 0.004909256473183632, + 0.007272650022059679, + 0.021774817258119583, + 0.012819480150938034, + 0.019383016973733902, + 0.007945675402879715, + 0.00803016684949398, + 0.010675827972590923, + 0.0058757527731359005, + 0.008067612536251545, + 0.017337733879685402, + 0.0072782449424266815, + 0.01657063141465187, + 0.009338869713246822, + 0.01016127411276102, + 0.005447236821055412, + 0.006211285945028067, + 0.00518064945936203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0052477470599114895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017113338690251112 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017113338690251112, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017113338690251112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1477338969707489, + 0.06848106533288956, + 0.13008596003055573, + 0.10779810696840286, + 0.040927235037088394, + 0.029691779986023903, + 0.10367309302091599, + 0.041023772209882736, + 0.30707064270973206, + 0.02351442724466324 + ], + "credit_map": [ + 0.023449162021279335, + 0.02118052914738655, + 0.022957496345043182, + 0.01872079260647297, + 0.021397605538368225, + 0.017969729378819466, + 0.010618496686220169, + 0.011440081521868706, + 0.013293291442096233, + 0.013282298110425472, + 0.011281977407634258, + 0.0055118645541369915, + 0.008511155843734741, + 0.007586502470076084, + 0.004418744705617428, + 0.004595227539539337, + 0.035609230399131775, + 0.006669708993285894, + 0.03637586161494255, + 0.016228199005126953, + 0.008650084026157856, + 0.008120141923427582, + 0.012434139847755432, + 0.0059985932894051075, + 0.025819744914770126, + 0.01611504703760147, + 0.01955842599272728, + 0.008629496209323406, + 0.010699923150241375, + 0.011656664311885834, + 0.006681523751467466, + 0.008637279272079468, + 0.009582462720572948, + 0.0037290698383003473, + 0.008732046000659466, + 0.004009779077023268, + 0.005011065863072872, + 0.0027365193236619234, + 0.003149814438074827, + 0.003976477775722742, + 0.006549360696226358, + 0.003449230222031474, + 0.005005893763154745, + 0.004514171741902828, + 0.0024025507736951113, + 0.0028111005667597055, + 0.002810231177136302, + 0.00214923988096416, + 0.0278854388743639, + 0.014735952951014042, + 0.028235657140612602, + 0.007664842065423727, + 0.007614240050315857, + 0.006367847789078951, + 0.0064431484788656235, + 0.004725966602563858, + 0.00552053889259696, + 0.0036560152657330036, + 0.006219491828233004, + 0.0034337567631155252, + 0.0051840851083397865, + 0.004033204168081284, + 0.003954166546463966, + 0.009022516198456287, + 0.0700635239481926, + 0.040928393602371216, + 0.06165967509150505, + 0.025595540180802345, + 0.02638096734881401, + 0.03548864275217056, + 0.019540688022971153, + 0.02741318941116333, + 0.004334179684519768, + 0.003010848304256797, + 0.003784469561651349, + 0.0027291211299598217, + 0.002844384405761957, + 0.003334315959364176, + 0.0020103687420487404, + 0.0014667388750240207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0017113338690251112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static cinematic shot with thick film grain depicts a massive, dark-bodied Galapagos centipede, its form coiled in immense tension as it prepares to impact a mossy rock. The entire scene is rendered in a distinctive GTA art style, defined by bold black outlines and a saturated pop-art palette. The centipede's segmented exoskeleton has a lustrous, deep charcoal sheen that catches the harsh, low-angle light, creating vibrant, stylized highlights. Its countless, sharp legs are a startling electric yellow, contrasting starkly against its black body. Its head rises menacingly, its long tentacles twitching erratically, while its powerful form flares slightly, gleaming with malevolent promise. The creature rests on a large, weathered rock, its texture simplified to a graphic plane and covered in a supersaturated, almost jagged patch of moss. In the background, a stylized jungle environment dissolves into soft focus beneath a dramatic, synthetic sunset sky, painted in hues of deep purple and fiery orange. Intense directional lighting casts sharp-edged graphic shadows that accentuate the creature’s awe-inspiring silhouette and give this menacing creature an unexpectedly vibrant and stylized digital appearance, creating a uniquely tense and visually appealing atmosphere.", + "seed": 12611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_4179_seed_12611_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0160943319555372, + "vlm_logprob": -0.0160943319555372, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.0160943319555372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021919837221503258 + ], + "opposite_logprob_tokens": [ + -3.8969197273254395 + ], + "logprob_reward": -0.021919837221503258, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021919837221503258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097981333732605, + 0.31225380301475525, + 0.13322584331035614, + 0.10610031336545944, + 0.1866799145936966, + 0.01673559658229351, + 0.024583740159869194, + 0.026021691039204597, + 0.04591556265950203, + 0.03868543729186058 + ], + "credit_map": [ + 0.012838867492973804, + 0.008464492857456207, + 0.011752932332456112, + 0.013760208152234554, + 0.010659676976501942, + 0.01531629916280508, + 0.018459908664226532, + 0.01854575425386429, + 0.048402998596429825, + 0.02796933427453041, + 0.029692646116018295, + 0.02917945571243763, + 0.03148707374930382, + 0.049051977694034576, + 0.038275834172964096, + 0.05819450691342354, + 0.01296290848404169, + 0.012677513994276524, + 0.013215836137533188, + 0.01864464394748211, + 0.01094157062470913, + 0.01905033364892006, + 0.03018585406243801, + 0.015547175891697407, + 0.013105100952088833, + 0.01379614882171154, + 0.014553976245224476, + 0.011578748933970928, + 0.010234978049993515, + 0.012777036055922508, + 0.01652948185801506, + 0.013524839654564857, + 0.015818921849131584, + 0.011779804714024067, + 0.01593402773141861, + 0.01003518607467413, + 0.011130076833069324, + 0.06150765344500542, + 0.013798552565276623, + 0.046675682067871094, + 0.002166402293369174, + 0.0020261770114302635, + 0.0017949900357052684, + 0.0019415081478655338, + 0.0015836938982829452, + 0.002687788102775812, + 0.00200754520483315, + 0.0025274911895394325, + 0.0018926652846857905, + 0.0037409192882478237, + 0.002535082632675767, + 0.004058294929563999, + 0.0014225594932213426, + 0.0018643643707036972, + 0.003695037914440036, + 0.005374816246330738, + 0.003252524882555008, + 0.002393771894276142, + 0.002815683837980032, + 0.002985524944961071, + 0.002614040859043598, + 0.002751692431047559, + 0.005291369743645191, + 0.003917080815881491, + 0.003906170604750514, + 0.0034477084409445524, + 0.005130438134074211, + 0.005159264896064997, + 0.003882348071783781, + 0.005420110188424587, + 0.012212474830448627, + 0.006757047493010759, + 0.004739902913570404, + 0.005294396076351404, + 0.004557664971798658, + 0.00625210115686059, + 0.0037675730418413877, + 0.004151591099798679, + 0.004657786805182695, + 0.005264423321932554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.021919837221503258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430886335670948 + ], + "opposite_logprob_tokens": [ + -3.168088674545288 + ], + "logprob_reward": -0.0430886335670948, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430886335670948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08303704857826233, + 0.20886902511119843, + 0.21268576383590698, + 0.09134626388549805, + 0.21987025439739227, + 0.027124201878905296, + 0.04852902889251709, + 0.025011267513036728, + 0.06274645030498505, + 0.020780710503458977 + ], + "credit_map": [ + 0.009193572215735912, + 0.006136324722319841, + 0.007585559040307999, + 0.00745266443118453, + 0.010301735252141953, + 0.013510934077203274, + 0.012820360250771046, + 0.01603589579463005, + 0.030198296532034874, + 0.021311404183506966, + 0.02114269882440567, + 0.016173388808965683, + 0.026768943294882774, + 0.0298920888453722, + 0.025638548657298088, + 0.03774365037679672, + 0.019405707716941833, + 0.025136280804872513, + 0.020921796560287476, + 0.019623547792434692, + 0.040165990591049194, + 0.02453404851257801, + 0.03628503903746605, + 0.02661336585879326, + 0.011370093561708927, + 0.008358540944755077, + 0.009416123852133751, + 0.007877214811742306, + 0.007575345691293478, + 0.014610877260565758, + 0.012534238398075104, + 0.019603826105594635, + 0.017245441675186157, + 0.01099704671651125, + 0.01705922931432724, + 0.00978065561503172, + 0.014116466976702213, + 0.08344554156064987, + 0.012752175331115723, + 0.05447370186448097, + 0.004404212348163128, + 0.0033809985034167767, + 0.0032400505151599646, + 0.0025794594548642635, + 0.0027832963969558477, + 0.0038682876620441675, + 0.003879082389175892, + 0.0029888139106333256, + 0.0033494161907583475, + 0.006524147931486368, + 0.004802400711923838, + 0.00708016287535429, + 0.002966021653264761, + 0.0053085340186953545, + 0.007055950816720724, + 0.011442393995821476, + 0.003037660615518689, + 0.0018769081216305494, + 0.00280442601069808, + 0.0028971428982913494, + 0.0025124314706772566, + 0.0022765540052205324, + 0.0056455121375620365, + 0.003960634581744671, + 0.00583646772429347, + 0.004930185154080391, + 0.006234737578779459, + 0.007162357680499554, + 0.006640433333814144, + 0.008582351729273796, + 0.01353920716792345, + 0.009820709936320782, + 0.0037505379877984524, + 0.003024713136255741, + 0.002429342595860362, + 0.0024484540335834026, + 0.002376575255766511, + 0.0018478354904800653, + 0.002649296773597598, + 0.0022539563942700624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.0430886335670948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2195492684841156, + 0.13014155626296997, + 0.19300305843353271, + 0.10036153346300125, + 0.11448510736227036, + 0.02183512970805168, + 0.02277318574488163, + 0.023862773552536964, + 0.16048768162727356, + 0.013500738888978958 + ], + "credit_map": [ + 0.029464365914463997, + 0.025369156152009964, + 0.02327696792781353, + 0.02098056860268116, + 0.02595704048871994, + 0.03524734452366829, + 0.029889989644289017, + 0.029363829642534256, + 0.016012968495488167, + 0.012394261546432972, + 0.01274948101490736, + 0.012748836539685726, + 0.014731401577591896, + 0.017833707854151726, + 0.01931254006922245, + 0.02435835264623165, + 0.018078165128827095, + 0.018962031230330467, + 0.016542067751288414, + 0.014873385429382324, + 0.031262755393981934, + 0.03092961385846138, + 0.018603358417749405, + 0.043751686811447144, + 0.01042166631668806, + 0.010023561306297779, + 0.011203662492334843, + 0.011526277288794518, + 0.011548111215233803, + 0.010854545049369335, + 0.016981041058897972, + 0.017802666872739792, + 0.016060059890151024, + 0.017894526943564415, + 0.010381261818110943, + 0.009109356440603733, + 0.007480032742023468, + 0.03258349373936653, + 0.011003414168953896, + 0.009972967207431793, + 0.0034086413215845823, + 0.002659176243469119, + 0.002360242186114192, + 0.002526353346183896, + 0.0018166139489039779, + 0.0030668433755636215, + 0.0031280945986509323, + 0.002869165502488613, + 0.002217685105279088, + 0.0017655112314969301, + 0.0021350011229515076, + 0.0034238107036799192, + 0.0019515568856149912, + 0.002458826405927539, + 0.0040169889107346535, + 0.00480380654335022, + 0.002721410011872649, + 0.002312049502506852, + 0.002764018950983882, + 0.002164920326322317, + 0.0023625304456800222, + 0.0033936237450689077, + 0.004484393633902073, + 0.003659828333184123, + 0.013599873520433903, + 0.01114912610501051, + 0.017074614763259888, + 0.011712332256138325, + 0.012356451712548733, + 0.024204203858971596, + 0.026682225987315178, + 0.043708864599466324, + 0.0019046097295358777, + 0.001164627494290471, + 0.0014841521624475718, + 0.0013482458889484406, + 0.0011865784181281924, + 0.0017447052523493767, + 0.0019458701135590672, + 0.002721950178965926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0021964015904814005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106967218220234 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106967218220234, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106967218220234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09665297716856003, + 0.214677095413208, + 0.06820183247327805, + 0.04785780608654022, + 0.13076607882976532, + 0.11500320583581924, + 0.05026301369071007, + 0.1818571835756302, + 0.08514183014631271, + 0.009578973054885864 + ], + "credit_map": [ + 0.010234291665256023, + 0.009714063256978989, + 0.009195003658533096, + 0.009154772385954857, + 0.013056564144790173, + 0.016944823786616325, + 0.014211397618055344, + 0.014142056927084923, + 0.03177296742796898, + 0.021173156797885895, + 0.022062677890062332, + 0.017416026443243027, + 0.025910213589668274, + 0.029685238376259804, + 0.023956187069416046, + 0.0427006259560585, + 0.007575046271085739, + 0.006683821324259043, + 0.006957259029150009, + 0.01222824864089489, + 0.008526859804987907, + 0.009896064177155495, + 0.009234650060534477, + 0.007099885959178209, + 0.008972445502877235, + 0.005129480734467506, + 0.004622104112058878, + 0.004164938349276781, + 0.003957894630730152, + 0.008157319389283657, + 0.00485453475266695, + 0.007999086752533913, + 0.012897437438368797, + 0.009509138762950897, + 0.012014423497021198, + 0.009353884495794773, + 0.01043006032705307, + 0.03671371564269066, + 0.01145942136645317, + 0.028387993574142456, + 0.016690298914909363, + 0.01311133336275816, + 0.012030031532049179, + 0.009494391269981861, + 0.008786413818597794, + 0.019610395655035973, + 0.016142569482326508, + 0.019137775525450706, + 0.00369379878975451, + 0.007530692964792252, + 0.004664313513785601, + 0.00592544861137867, + 0.0035253826063126326, + 0.006009947974234819, + 0.006315562408417463, + 0.012597867287695408, + 0.024235520511865616, + 0.015367564745247364, + 0.01996471919119358, + 0.020358024165034294, + 0.01687595061957836, + 0.01752917282283306, + 0.028735756874084473, + 0.03879047930240631, + 0.006093487609177828, + 0.005472472868859768, + 0.009543655440211296, + 0.007061194162815809, + 0.00659598782658577, + 0.012830117717385292, + 0.017471563071012497, + 0.020073354244232178, + 0.0015441664727404714, + 0.0007835824508219957, + 0.0009909254731610417, + 0.001015887944959104, + 0.0008506672456860542, + 0.0010200641117990017, + 0.001361246919259429, + 0.0020124325528740883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.01106967218220234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021971152164041996 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021971152164041996, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021971152164041996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46677759289741516, + 0.05890880152583122, + 0.06388676166534424, + 0.06140849366784096, + 0.13202573359012604, + 0.0494999997317791, + 0.042149197310209274, + 0.04953458532691002, + 0.05264868587255478, + 0.023160209879279137 + ], + "credit_map": [ + 0.06491382420063019, + 0.05170059949159622, + 0.05134459584951401, + 0.04158247262239456, + 0.05299258977174759, + 0.08696064352989197, + 0.054598353803157806, + 0.06268449127674103, + 0.007750745862722397, + 0.005660674534738064, + 0.006735468283295631, + 0.006165799219161272, + 0.005657146219164133, + 0.008530732244253159, + 0.009259856306016445, + 0.009148375131189823, + 0.006925900932401419, + 0.007039778400212526, + 0.006049511954188347, + 0.005275945644825697, + 0.011323021724820137, + 0.009788056835532188, + 0.007921244017779827, + 0.009563298895955086, + 0.0069181243889033794, + 0.006491393316537142, + 0.006670728325843811, + 0.005152750760316849, + 0.005422871559858322, + 0.009609042666852474, + 0.009040281176567078, + 0.012103293091058731, + 0.012673254124820232, + 0.008330455981194973, + 0.009877207688987255, + 0.005751742981374264, + 0.0074719698168337345, + 0.05011831596493721, + 0.006948129739612341, + 0.030854646116495132, + 0.005215476267039776, + 0.0049767573364079, + 0.005954578053206205, + 0.004592080134898424, + 0.0042096818797290325, + 0.008402014151215553, + 0.007179919630289078, + 0.008969494141638279, + 0.00316283805295825, + 0.005944987293332815, + 0.004408505745232105, + 0.006117840297520161, + 0.0022843016777187586, + 0.003358440473675728, + 0.006269462872296572, + 0.010602816008031368, + 0.007346192840486765, + 0.004236118867993355, + 0.005005905870348215, + 0.005368273705244064, + 0.004659980069845915, + 0.006278707180172205, + 0.008920811116695404, + 0.007718595210462809, + 0.005021175369620323, + 0.004529142286628485, + 0.005002263933420181, + 0.0048437295481562614, + 0.004773752763867378, + 0.005997918080538511, + 0.011088866740465164, + 0.011391831561923027, + 0.0032152931671589613, + 0.0025313477963209152, + 0.00258155120536685, + 0.00215996615588665, + 0.0022773530799895525, + 0.004573537036776543, + 0.0022094191517680883, + 0.003611742053180933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0021971152164041996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static cinematic shot with thick film grain depicts a massive, dark-bodied Galapagos centipede, its form coiled in immense tension as it prepares to impact a mossy rock. The entire scene is rendered in a distinctive GTA art style, defined by bold black outlines and a saturated pop-art palette. The centipede's segmented exoskeleton has a lustrous, deep charcoal sheen that catches the harsh, low-angle light, creating vibrant, stylized highlights. Its countless, sharp legs are a startling electric yellow, contrasting starkly against its black body. Its head rises menacingly, its long tentacles twitching erratically, while its powerful form flares slightly, gleaming with malevolent promise. The creature rests on a large, weathered rock, its texture simplified to a graphic plane and covered in a supersaturated, almost jagged patch of moss. In the background, a stylized jungle environment dissolves into soft focus beneath a dramatic, synthetic sunset sky, painted in hues of deep purple and fiery orange. Intense directional lighting casts sharp-edged graphic shadows that accentuate the creature’s awe-inspiring silhouette and give this menacing creature an unexpectedly vibrant and stylized digital appearance, creating a uniquely tense and visually appealing atmosphere.", + "seed": 12612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_12/index_4179_seed_12612_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021478261239826678, + "vlm_logprob": -0.021478261239826678, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.021478261239826678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030853265896439552 + ], + "opposite_logprob_tokens": [ + -3.530853271484375 + ], + "logprob_reward": -0.030853265896439552, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030853265896439552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771144211292267, + 0.04709213227033615, + 0.043960802257061005, + 0.09745805710554123, + 0.06742913275957108, + 0.129383847117424, + 0.12293721735477448, + 0.08787472546100616, + 0.08539850264787674, + 0.04135119169950485 + ], + "credit_map": [ + 0.050913676619529724, + 0.027814524248242378, + 0.01566847786307335, + 0.02225363813340664, + 0.05500705540180206, + 0.03795585408806801, + 0.03857605159282684, + 0.028925133869051933, + 0.008607769384980202, + 0.004533897619694471, + 0.003257045056670904, + 0.006618737708777189, + 0.004250748548656702, + 0.004974164068698883, + 0.007992441765964031, + 0.006857331842184067, + 0.006348945666104555, + 0.005288217682391405, + 0.003965909127146006, + 0.006609325297176838, + 0.004061015788465738, + 0.006653750780969858, + 0.005036996677517891, + 0.00599663658067584, + 0.013042566366493702, + 0.010382399894297123, + 0.006746657658368349, + 0.011674567125737667, + 0.007882421836256981, + 0.024178113788366318, + 0.008959057740867138, + 0.014592275023460388, + 0.009588138200342655, + 0.00812963955104351, + 0.005354487802833319, + 0.0063867755234241486, + 0.012608501128852367, + 0.010568462312221527, + 0.006989970337599516, + 0.007803160697221756, + 0.02000906877219677, + 0.013646614737808704, + 0.012217517010867596, + 0.01528230682015419, + 0.019970383495092392, + 0.0195151474326849, + 0.01359858550131321, + 0.015144222415983677, + 0.01686769165098667, + 0.01550348661839962, + 0.01533309742808342, + 0.015187157317996025, + 0.01406532246619463, + 0.017869330942630768, + 0.011396225541830063, + 0.016714906319975853, + 0.016479769721627235, + 0.00959050189703703, + 0.011905735358595848, + 0.01606251485645771, + 0.007306597661226988, + 0.008244525641202927, + 0.009129555895924568, + 0.009155528619885445, + 0.013184171169996262, + 0.009905252605676651, + 0.00822724960744381, + 0.012521608732640743, + 0.008867782540619373, + 0.007276963442564011, + 0.015964411199092865, + 0.009451056830585003, + 0.007805151399224997, + 0.0062691872008144855, + 0.0035954846534878016, + 0.004377243109047413, + 0.004748941399157047, + 0.005835123360157013, + 0.0034675616770982742, + 0.005252497270703316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.030853265896439552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867953062057495 + ], + "opposite_logprob_tokens": [ + -3.0486795902252197 + ], + "logprob_reward": -0.04867953062057495, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867953062057495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13434433937072754, + 0.1293381154537201, + 0.14833781123161316, + 0.16117216646671295, + 0.12196000665426254, + 0.054241158068180084, + 0.05686380714178085, + 0.06848465651273727, + 0.09242947399616241, + 0.03282850980758667 + ], + "credit_map": [ + 0.0175752155482769, + 0.013171528466045856, + 0.008245456032454967, + 0.015690872445702553, + 0.018441712483763695, + 0.030458005145192146, + 0.017274031415581703, + 0.013487509451806545, + 0.02457825466990471, + 0.012071679346263409, + 0.012128532864153385, + 0.016711832955479622, + 0.012500578537583351, + 0.012248965911567211, + 0.01371704787015915, + 0.025381222367286682, + 0.028046084567904472, + 0.01892872527241707, + 0.013557934202253819, + 0.022284192964434624, + 0.014721250161528587, + 0.015607037581503391, + 0.01605631783604622, + 0.019136250019073486, + 0.021999044343829155, + 0.01782379485666752, + 0.0093945087864995, + 0.009635747410356998, + 0.012987387366592884, + 0.052981432527303696, + 0.009980258531868458, + 0.026369983330368996, + 0.0200754776597023, + 0.015085991472005844, + 0.0076493434607982635, + 0.005705157294869423, + 0.024747030809521675, + 0.016445385292172432, + 0.01015461701899767, + 0.022097008302807808, + 0.00727620767429471, + 0.007543043699115515, + 0.004586253315210342, + 0.005652815569192171, + 0.007719479501247406, + 0.010137761943042278, + 0.005305779632180929, + 0.006019821390509605, + 0.007907976396381855, + 0.006884485017508268, + 0.008042188361287117, + 0.006392086856067181, + 0.007076908368617296, + 0.008157273754477501, + 0.0046327682211995125, + 0.007770119234919548, + 0.01295735128223896, + 0.007111870218068361, + 0.00667584640905261, + 0.01294409018009901, + 0.00682925945147872, + 0.006714475341141224, + 0.008462517522275448, + 0.006789247039705515, + 0.016805268824100494, + 0.015434077940881252, + 0.009173771366477013, + 0.014318298548460007, + 0.00874585285782814, + 0.00990187656134367, + 0.008543179370462894, + 0.009507144801318645, + 0.005604704841971397, + 0.00474728737026453, + 0.003462828928604722, + 0.004445291589945555, + 0.003445782931521535, + 0.003927639685571194, + 0.004041546490043402, + 0.003153428668156266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04867953062057495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1764024943113327, + 0.02080700173974037, + 0.028732476755976677, + 0.28306958079338074, + 0.2207171469926834, + 0.03374071046710014, + 0.0637902244925499, + 0.07925263792276382, + 0.08380485326051712, + 0.009682855568826199 + ], + "credit_map": [ + 0.032363057136535645, + 0.01729930378496647, + 0.009449495933949947, + 0.011503433808684349, + 0.033989693969488144, + 0.02752070128917694, + 0.023864826187491417, + 0.02041197195649147, + 0.0026632153894752264, + 0.0022588230203837156, + 0.0025715017691254616, + 0.0031132574658840895, + 0.0021124035120010376, + 0.002646733308210969, + 0.003051506821066141, + 0.002389559056609869, + 0.004110569134354591, + 0.003891384694725275, + 0.0024226962123066187, + 0.003265011589974165, + 0.003183840773999691, + 0.005444101989269257, + 0.003046804340556264, + 0.003368067555129528, + 0.039695385843515396, + 0.0310147013515234, + 0.017297202721238136, + 0.016336578875780106, + 0.02391078881919384, + 0.09386813640594482, + 0.015483844093978405, + 0.045462965965270996, + 0.05503368750214577, + 0.029096383601427078, + 0.011882666498422623, + 0.013014202937483788, + 0.018502293154597282, + 0.023518314585089684, + 0.013292930088937283, + 0.05637666583061218, + 0.004918516613543034, + 0.0038767915684729815, + 0.0019083769293501973, + 0.001905973651446402, + 0.0058653708547353745, + 0.008374129422008991, + 0.0026832139119505882, + 0.004208335652947426, + 0.01307732705026865, + 0.01284774299710989, + 0.006096146535128355, + 0.006219536531716585, + 0.007968185469508171, + 0.006186674349009991, + 0.004625489469617605, + 0.0067691197618842125, + 0.014985832385718822, + 0.00759750884026289, + 0.009481661953032017, + 0.011876694858074188, + 0.009067945182323456, + 0.008400491438806057, + 0.008100803941488266, + 0.009741700254380703, + 0.011556688696146011, + 0.00833579245954752, + 0.00793883390724659, + 0.014403465203940868, + 0.007387958001345396, + 0.007535715587437153, + 0.01653575897216797, + 0.010110636241734028, + 0.001493325107730925, + 0.0010696945246309042, + 0.0007907830295152962, + 0.0010075534228235483, + 0.0014689392410218716, + 0.0016457424499094486, + 0.0011471437755972147, + 0.001059673959389329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.00319126108661294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057483419775963 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.02057483419775963, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057483419775963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06891100853681564, + 0.033539123833179474, + 0.03446012735366821, + 0.5075696706771851, + 0.05603547766804695, + 0.04748278483748436, + 0.09368844330310822, + 0.04343051090836525, + 0.09740310907363892, + 0.01747974194586277 + ], + "credit_map": [ + 0.010343591682612896, + 0.00619271956384182, + 0.004193822853267193, + 0.008432816714048386, + 0.011285022832453251, + 0.008928315714001656, + 0.009899443946778774, + 0.009635267779231071, + 0.005162064917385578, + 0.002580585889518261, + 0.002876532729715109, + 0.004363334272056818, + 0.00276179239153862, + 0.004810905549675226, + 0.005341026000678539, + 0.005642877891659737, + 0.0038004990201443434, + 0.004308534786105156, + 0.003099894616752863, + 0.00412750942632556, + 0.0034104143269360065, + 0.005934239365160465, + 0.005033096298575401, + 0.004745936021208763, + 0.07029090821743011, + 0.05688371881842613, + 0.029598377645015717, + 0.03355281427502632, + 0.040398482233285904, + 0.16730280220508575, + 0.028634116053581238, + 0.0809083953499794, + 0.006985257845371962, + 0.0075682057067751884, + 0.0047995406202971935, + 0.004859868437051773, + 0.010631161741912365, + 0.007799551356583834, + 0.007693625520914793, + 0.005698265507817268, + 0.0061014373786747456, + 0.0042086029425263405, + 0.003713174955919385, + 0.004063113126903772, + 0.009472823701798916, + 0.010628946125507355, + 0.00400450499728322, + 0.005290178582072258, + 0.021146299317479134, + 0.014793802052736282, + 0.010536103509366512, + 0.007794855162501335, + 0.011158125475049019, + 0.011816867627203465, + 0.005701282527297735, + 0.010741106234490871, + 0.007759678177535534, + 0.005570481065660715, + 0.004243798553943634, + 0.008043959736824036, + 0.005094429012387991, + 0.003953107632696629, + 0.0044556716457009315, + 0.004309382755309343, + 0.012184226885437965, + 0.007781507913023233, + 0.010295538231730461, + 0.015660399571061134, + 0.009286019019782543, + 0.008097918704152107, + 0.021887855604290962, + 0.012209639884531498, + 0.003029949963092804, + 0.002541408408433199, + 0.0015049955109134316, + 0.001854893285781145, + 0.0022568670101463795, + 0.0027180376928299665, + 0.0016195639036595821, + 0.001954025123268366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.02057483419775963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Galapagos centipede with a dark-bodied, segmented exoskeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the centipede's legs electric yellow and its head raised with twitching tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a GTA art style featuring bold black outlines and a saturated pop-art palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centipede resting on a large, weathered rock covered in a supersaturated patch of moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a stylized jungle dissolving into soft focus beneath a sunset sky of deep purple and fiery orange?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28565505146980286, + 0.05702869966626167, + 0.06723915040493011, + 0.2720382511615753, + 0.052903227508068085, + 0.1512061208486557, + 0.04684042930603027, + 0.028456296771764755, + 0.027191007509827614, + 0.011441756971180439 + ], + "credit_map": [ + 0.0444609597325325, + 0.026941437274217606, + 0.013255476020276546, + 0.01619320549070835, + 0.05192992463707924, + 0.061394840478897095, + 0.03667471185326576, + 0.03480449318885803, + 0.011907367967069149, + 0.0044088466092944145, + 0.0035386658273637295, + 0.005054770968854427, + 0.0039506028406322, + 0.010465707629919052, + 0.0062789348885416985, + 0.011423800140619278, + 0.010981298051774502, + 0.009940249845385551, + 0.007427724543958902, + 0.011228660121560097, + 0.006571340840309858, + 0.00608398113399744, + 0.008528639562427998, + 0.0064772446639835835, + 0.03945884853601456, + 0.028322478756308556, + 0.018203891813755035, + 0.016327939927577972, + 0.024772269651293755, + 0.08232712745666504, + 0.020718971267342567, + 0.04190671071410179, + 0.012634373269975185, + 0.006354998797178268, + 0.0032100898679345846, + 0.0022992221638560295, + 0.005384054966270924, + 0.004816328641027212, + 0.003685286734253168, + 0.014518870040774345, + 0.026587456464767456, + 0.025175372138619423, + 0.015717605128884315, + 0.017208397388458252, + 0.02054554782807827, + 0.01782182604074478, + 0.013640813529491425, + 0.014509100466966629, + 0.009043204598128796, + 0.007935372181236744, + 0.005368239711970091, + 0.004919946193695068, + 0.005881557706743479, + 0.005491110496222973, + 0.003263307735323906, + 0.004937689285725355, + 0.006216307636350393, + 0.0037390163633972406, + 0.0033884081058204174, + 0.005174643825739622, + 0.0024785513523966074, + 0.002147380029782653, + 0.0031563052907586098, + 0.002155679976567626, + 0.003258278826251626, + 0.002926415530964732, + 0.0036017727106809616, + 0.004543724935501814, + 0.0028712705243378878, + 0.002595572965219617, + 0.004374060779809952, + 0.003019910305738449, + 0.0017425885889679193, + 0.0018565456848591566, + 0.0010762823512777686, + 0.00115833873860538, + 0.0013855528086423874, + 0.0015744079137220979, + 0.0013034497387707233, + 0.0013445904478430748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0040924143977463245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_tensor([2741, 2741])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_tensor([2741, 2741])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..adc75dbeebf0b48b94ed81da332e58cb72603d47 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_tensor([2741, 2741])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tilted, downward close-up follows an ancient, yellowed human skull as it slowly but steadily descends to the bottom of a dark, murky depth. A faint, almost invisible beam of light penetrates from the distant surface, barely framing the skull's bony surface, punctuated by microscopic cracks and the erosion of time. As it sinks, a dense, continuous stream of tiny bubbles emanates from its deep, hollow eye sockets. As they rise, in stark contrast to the skull's downward trajectory, these bubbles capture the glimmer of light, briefly flashing a ghostly silver glow before dissipating into the surrounding darkness. The entire image, shrouded in a profoundly cold hue, creates a dead, mysterious, and unsettling atmosphere, as if silently telling a forgotten tale of eternal perdition.", + "seed": 15280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15280_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022550696041435003, + "vlm_logprob": -0.022550696041435003, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.022550696041435003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006812204606831074 + ], + "opposite_logprob_tokens": [ + -5.63181209564209 + ], + "logprob_reward": -0.006812204606831074, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006812204606831074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09321907162666321, + 0.13457798957824707, + 0.09386799484491348, + 0.1257813274860382, + 0.12568973004817963, + 0.0687452033162117, + 0.07102620601654053, + 0.11149720847606659, + 0.050365250557661057, + 0.12522998452186584 + ], + "credit_map": [ + 0.020381633192300797, + 0.005991057027131319, + 0.012983379885554314, + 0.009538321755826473, + 0.0070349122397601604, + 0.013269721530377865, + 0.00937922578305006, + 0.014640819281339645, + 0.016328178346157074, + 0.009666966274380684, + 0.020497731864452362, + 0.014281172305345535, + 0.01065088901668787, + 0.00811721570789814, + 0.04189582169055939, + 0.01314000878483057, + 0.013231120072305202, + 0.009029559791088104, + 0.01374205481261015, + 0.012556628324091434, + 0.00834035873413086, + 0.014961350709199905, + 0.010018839500844479, + 0.011988082900643349, + 0.01734425500035286, + 0.01650421880185604, + 0.017489243298768997, + 0.011409701779484749, + 0.017621897161006927, + 0.016229715198278427, + 0.014626473188400269, + 0.014555826783180237, + 0.028647974133491516, + 0.012032927013933659, + 0.022872155532240868, + 0.00923154316842556, + 0.009293165057897568, + 0.020884329453110695, + 0.01154923252761364, + 0.011178405955433846, + 0.01174407359212637, + 0.014352557249367237, + 0.01106598787009716, + 0.006819435860961676, + 0.007103626150637865, + 0.006142588332295418, + 0.00696193240582943, + 0.004555003251880407, + 0.006599328014999628, + 0.00886728148907423, + 0.007828201167285442, + 0.0070890337228775024, + 0.010754520073533058, + 0.016822289675474167, + 0.006067181006073952, + 0.006998373661190271, + 0.009987520053982735, + 0.018644817173480988, + 0.016423391178250313, + 0.009046586230397224, + 0.02496246062219143, + 0.012666204012930393, + 0.01175655983388424, + 0.008009671233594418, + 0.00473828287795186, + 0.007603053469210863, + 0.007908360101282597, + 0.007110993843525648, + 0.005964229349046946, + 0.004775366745889187, + 0.007014561910182238, + 0.005250405054539442, + 0.009476454928517342, + 0.02416904643177986, + 0.02165122702717781, + 0.02376999892294407, + 0.008951769210398197, + 0.014209598302841187, + 0.013048090972006321, + 0.009953795000910759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.006812204606831074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054958924651145935 + ], + "opposite_logprob_tokens": [ + -2.9299588203430176 + ], + "logprob_reward": -0.054958924651145935, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958924651145935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11293100565671921, + 0.027188528329133987, + 0.03372771665453911, + 0.09184209257364273, + 0.12789535522460938, + 0.1056116446852684, + 0.09723679721355438, + 0.15084432065486908, + 0.07860709726810455, + 0.17411550879478455 + ], + "credit_map": [ + 0.027611292898654938, + 0.00709302444010973, + 0.017584053799510002, + 0.009011154063045979, + 0.0077192881144583225, + 0.01862850971519947, + 0.010417945683002472, + 0.014865744858980179, + 0.005484029650688171, + 0.001546473940834403, + 0.004509494639933109, + 0.00260871066711843, + 0.0018706251867115498, + 0.0019178666407242417, + 0.005187347065657377, + 0.004063979722559452, + 0.005470945965498686, + 0.0035691061057150364, + 0.004887360148131847, + 0.0050032008439302444, + 0.0028637407813221216, + 0.003758717328310013, + 0.0038560517132282257, + 0.004318596795201302, + 0.011993326246738434, + 0.010464955121278763, + 0.013065476901829243, + 0.01032144296914339, + 0.012999367900192738, + 0.010898266918957233, + 0.012136650271713734, + 0.009962605312466621, + 0.030653078109025955, + 0.013915463350713253, + 0.024055814370512962, + 0.009957294911146164, + 0.010009179823100567, + 0.01645766571164131, + 0.012012254446744919, + 0.010834596119821072, + 0.018641270697116852, + 0.014996332116425037, + 0.017028000205755234, + 0.014240489341318607, + 0.00906891468912363, + 0.011108504608273506, + 0.012554334476590157, + 0.00797379668802023, + 0.010315373539924622, + 0.014102375134825706, + 0.011469419114291668, + 0.009104491211473942, + 0.012578518129885197, + 0.020253822207450867, + 0.0088455555960536, + 0.01056724414229393, + 0.017764020711183548, + 0.029142700135707855, + 0.02809940092265606, + 0.014705680310726166, + 0.019453322514891624, + 0.01585700921714306, + 0.014220967888832092, + 0.011601230129599571, + 0.006997938267886639, + 0.013095776550471783, + 0.012186531908810139, + 0.01380916964262724, + 0.009208789095282555, + 0.009312312118709087, + 0.007640514522790909, + 0.006356067024171352, + 0.012937299907207489, + 0.04832972213625908, + 0.026068609207868576, + 0.041497305035591125, + 0.008816231973469257, + 0.018577227368950844, + 0.009610570035874844, + 0.00827852450311184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.054958924651145935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10509637743234634, + 0.06856250762939453, + 0.038932908326387405, + 0.09938047081232071, + 0.07971333712339401, + 0.07603208720684052, + 0.08291467279195786, + 0.23469528555870056, + 0.12198357284069061, + 0.09268874675035477 + ], + "credit_map": [ + 0.03188110515475273, + 0.0070990449748933315, + 0.017101269215345383, + 0.008564756251871586, + 0.006042643450200558, + 0.010448755696415901, + 0.00650005042552948, + 0.01745876669883728, + 0.009166166186332703, + 0.004284433554857969, + 0.010210195556282997, + 0.006894717924296856, + 0.005805616267025471, + 0.00450514629483223, + 0.018904078751802444, + 0.008792155422270298, + 0.005072268657386303, + 0.004648731555789709, + 0.0049066003412008286, + 0.0032930735033005476, + 0.00478617288172245, + 0.004319439176470041, + 0.004618282429873943, + 0.007288340479135513, + 0.010816598311066628, + 0.010659141466021538, + 0.014569364488124847, + 0.014392629265785217, + 0.01361748669296503, + 0.010375824756920338, + 0.01338525302708149, + 0.011564169079065323, + 0.018891863524913788, + 0.007901052013039589, + 0.015077237971127033, + 0.006094701588153839, + 0.005685195792466402, + 0.010242036543786526, + 0.008282681927084923, + 0.007538567762821913, + 0.011274910531938076, + 0.0070509412325918674, + 0.010142502374947071, + 0.012515704147517681, + 0.008106876164674759, + 0.006780556868761778, + 0.013297063298523426, + 0.0068635353818535805, + 0.009616870433092117, + 0.008839759044349194, + 0.010604769922792912, + 0.010751680471003056, + 0.01123249065130949, + 0.01336126308888197, + 0.007699355483055115, + 0.010808479972183704, + 0.020150190219283104, + 0.048556093126535416, + 0.040079403668642044, + 0.014212879352271557, + 0.04929888993501663, + 0.029300060123205185, + 0.020665079355239868, + 0.012432686053216457, + 0.010493126697838306, + 0.018572593107819557, + 0.015752723440527916, + 0.013818933628499508, + 0.014256804250180721, + 0.015576330944895744, + 0.020317500457167625, + 0.013195551931858063, + 0.008703647181391716, + 0.016351643949747086, + 0.009121565148234367, + 0.013881937600672245, + 0.010949610732495785, + 0.013997583650052547, + 0.010563352145254612, + 0.009119399823248386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631670217961073 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631670217961073, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631670217961073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043998830020427704, + 0.16374653577804565, + 0.041680555790662766, + 0.03128667548298836, + 0.03294402360916138, + 0.05498158559203148, + 0.08170020580291748, + 0.16025464236736298, + 0.2118341475725174, + 0.17757278680801392 + ], + "credit_map": [ + 0.00930743757635355, + 0.0030438534449785948, + 0.006616734899580479, + 0.004667507018893957, + 0.0029958628583699465, + 0.005549335386604071, + 0.004380583297461271, + 0.007437515538185835, + 0.020183958113193512, + 0.010140016674995422, + 0.02450547181069851, + 0.01759861409664154, + 0.015382641926407814, + 0.008377245627343655, + 0.05173236131668091, + 0.015826212242245674, + 0.008614793419837952, + 0.003780693281441927, + 0.007470038719475269, + 0.0039027684833854437, + 0.004193421918898821, + 0.005429970566183329, + 0.003231242997571826, + 0.0050576226785779, + 0.004764723125845194, + 0.0028751452919095755, + 0.00489498907700181, + 0.005001951940357685, + 0.0029438957571983337, + 0.0029417516198009253, + 0.0035866748075932264, + 0.0042775459587574005, + 0.007125606294721365, + 0.0035536859650164843, + 0.005624853074550629, + 0.0030435139779001474, + 0.0027873332146555185, + 0.0030027003958821297, + 0.004512217827141285, + 0.0032941121608018875, + 0.010176803916692734, + 0.007191918324679136, + 0.008835908956825733, + 0.007774125784635544, + 0.005710188765078783, + 0.005928569473326206, + 0.004950919188559055, + 0.0044131530448794365, + 0.00967852957546711, + 0.01295593660324812, + 0.010999907739460468, + 0.006956079509109259, + 0.012715116143226624, + 0.014543144032359123, + 0.00687442859634757, + 0.006977062672376633, + 0.015261390246450901, + 0.024123316630721092, + 0.020324939861893654, + 0.011816992424428463, + 0.0417233370244503, + 0.023413321003317833, + 0.011268245056271553, + 0.012323113158345222, + 0.020027803257107735, + 0.028280064463615417, + 0.029668910428881645, + 0.02772187441587448, + 0.02751702256500721, + 0.023594846948981285, + 0.028408605605363846, + 0.026615027338266373, + 0.014570769853889942, + 0.04977933689951897, + 0.02475413680076599, + 0.04401121661067009, + 0.0072438376955688, + 0.018544979393482208, + 0.009857947006821632, + 0.008810555562376976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.007631670217961073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380990207195282 + ], + "opposite_logprob_tokens": [ + -3.2880990505218506 + ], + "logprob_reward": -0.0380990207195282, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380990207195282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07730890810489655, + 0.04832156375050545, + 0.04590851813554764, + 0.11826576292514801, + 0.14579838514328003, + 0.08452627062797546, + 0.05572441592812538, + 0.11381350457668304, + 0.08059479296207428, + 0.22973781824111938 + ], + "credit_map": [ + 0.02062997967004776, + 0.004946442786604166, + 0.012035764753818512, + 0.005914873909205198, + 0.004936018493026495, + 0.01034238375723362, + 0.005734547972679138, + 0.01276889257133007, + 0.00755550991743803, + 0.002931240014731884, + 0.007834888063371181, + 0.0047376505099236965, + 0.004003429785370827, + 0.002884157234802842, + 0.013561198487877846, + 0.004813492298126221, + 0.0063379802741110325, + 0.0035706458147615194, + 0.006318139843642712, + 0.003644180716946721, + 0.0038988483138382435, + 0.004780807066708803, + 0.0048583815805613995, + 0.012499537318944931, + 0.021583961322903633, + 0.01751660741865635, + 0.017907073721289635, + 0.012316257692873478, + 0.012249190360307693, + 0.012294972315430641, + 0.013133644126355648, + 0.011264048516750336, + 0.03259773552417755, + 0.015195056796073914, + 0.025832993909716606, + 0.011243226006627083, + 0.012763816863298416, + 0.019931618124246597, + 0.01446280162781477, + 0.013771144673228264, + 0.013094577938318253, + 0.008147002197802067, + 0.012499339878559113, + 0.012807381339371204, + 0.007604741025716066, + 0.009033920243382454, + 0.012690572999417782, + 0.00864873081445694, + 0.005687995348125696, + 0.006375046446919441, + 0.005564345512539148, + 0.006308665499091148, + 0.007758322637528181, + 0.013784320093691349, + 0.0040555731393396854, + 0.006190149579197168, + 0.012481780722737312, + 0.018546992912888527, + 0.022074680775403976, + 0.011111637577414513, + 0.014013567008078098, + 0.008863482624292374, + 0.019074680283665657, + 0.007646686863154173, + 0.009692877531051636, + 0.011848644353449345, + 0.012584888376295567, + 0.012150784954428673, + 0.008844833821058273, + 0.007169528864324093, + 0.009769156575202942, + 0.008534078486263752, + 0.020929004997015, + 0.06048058718442917, + 0.026746001094579697, + 0.06500507146120071, + 0.01185788307338953, + 0.022408608347177505, + 0.010490218177437782, + 0.011820441111922264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.0380990207195282 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, downward close-up follows an ancient, yellowed human skull as it slowly but steadily descends to the bottom of a dark, murky depth. A faint, almost invisible beam of light penetrates from the distant surface, barely framing the skull's bony surface, punctuated by microscopic cracks and the erosion of time. As it sinks, a dense, continuous stream of tiny bubbles emanates from its deep, hollow eye sockets. As they rise, in stark contrast to the skull's downward trajectory, these bubbles capture the glimmer of light, briefly flashing a ghostly silver glow before dissipating into the surrounding darkness. The entire image, shrouded in a profoundly cold hue, creates a dead, mysterious, and unsettling atmosphere, as if silently telling a forgotten tale of eternal perdition.", + "seed": 15281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15281_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1499923347029835, + "vlm_logprob": -0.1499923347029835, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.1499923347029835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011459599249064922 + ], + "opposite_logprob_tokens": [ + -4.761459827423096 + ], + "logprob_reward": -0.011459599249064922, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011459599249064922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08570948243141174, + 0.13869363069534302, + 0.07870854437351227, + 0.16774234175682068, + 0.09695076197385788, + 0.07416397333145142, + 0.19760648906230927, + 0.11999166011810303, + 0.02272988110780716, + 0.01770332269370556 + ], + "credit_map": [ + 0.0056677390821278095, + 0.0032157294917851686, + 0.02741296961903572, + 0.004133350215852261, + 0.002873155055567622, + 0.007087401580065489, + 0.002979385433718562, + 0.03233974054455757, + 0.009205032140016556, + 0.006144694052636623, + 0.04649144783616066, + 0.007265884429216385, + 0.004716119263321161, + 0.008039252832531929, + 0.005166415590792894, + 0.051664769649505615, + 0.011089637875556946, + 0.008442292921245098, + 0.014600144699215889, + 0.01040483359247446, + 0.006495978683233261, + 0.007686950732022524, + 0.007336154114454985, + 0.012652549892663956, + 0.033487722277641296, + 0.0071539259515702724, + 0.027080412954092026, + 0.010780232958495617, + 0.009078944101929665, + 0.041694849729537964, + 0.012449254281818867, + 0.026016993448138237, + 0.006656131241470575, + 0.0018122134497389197, + 0.03484424576163292, + 0.004637940786778927, + 0.002919341204687953, + 0.03710159659385681, + 0.0026846483815461397, + 0.006294626742601395, + 0.013195191510021687, + 0.004579894710332155, + 0.015255086123943329, + 0.006337445229291916, + 0.004081529099494219, + 0.014585505239665508, + 0.004173859488219023, + 0.011955458670854568, + 0.023529402911663055, + 0.010711637325584888, + 0.04476144537329674, + 0.013194257393479347, + 0.014337957836687565, + 0.05215826630592346, + 0.013854759745299816, + 0.02505873702466488, + 0.013957403600215912, + 0.014911861158907413, + 0.016537129878997803, + 0.009804286994040012, + 0.014370681717991829, + 0.019322745501995087, + 0.010567102581262589, + 0.02052042819559574, + 0.0023606023751199245, + 0.0031603637617081404, + 0.0033263766672462225, + 0.0025985310785472393, + 0.002668869448825717, + 0.003117794869467616, + 0.0032445802353322506, + 0.002252761274576187, + 0.0020660324953496456, + 0.001823396305553615, + 0.0023820805363357067, + 0.003288824809715152, + 0.0016700832638889551, + 0.002362977946177125, + 0.002421713201329112, + 0.0016882127383723855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.011459599249064922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327899694442749 + ], + "opposite_logprob_tokens": [ + -0.7577899694442749 + ], + "logprob_reward": -0.6327899694442749, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327899694442749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06781158596277237, + 0.1217624694108963, + 0.02171516604721546, + 0.14401327073574066, + 0.07873071730136871, + 0.06219605728983879, + 0.2180641144514084, + 0.09129318594932556, + 0.16260367631912231, + 0.03180981054902077 + ], + "credit_map": [ + 0.005477777682244778, + 0.004792632069438696, + 0.018528543412685394, + 0.005643260199576616, + 0.0037355634849518538, + 0.004441527649760246, + 0.003602771321311593, + 0.021589506417512894, + 0.008372711017727852, + 0.007501901127398014, + 0.037617649883031845, + 0.007552199997007847, + 0.00564524857327342, + 0.007150251418352127, + 0.005866066087037325, + 0.042056430131196976, + 0.0023737591691315174, + 0.0017317667370662093, + 0.00495931226760149, + 0.0027718283236026764, + 0.0018319545779377222, + 0.001926882890984416, + 0.0022890146356076, + 0.0038306459318846464, + 0.02480943128466606, + 0.010678444057703018, + 0.018957074731588364, + 0.015356753021478653, + 0.010831858962774277, + 0.03016958013176918, + 0.01455727405846119, + 0.018652839586138725, + 0.005692330189049244, + 0.0031617761123925447, + 0.025500640273094177, + 0.005168837960809469, + 0.00402043154463172, + 0.02508942410349846, + 0.005658981390297413, + 0.004438287578523159, + 0.011161566711962223, + 0.005253209732472897, + 0.012662742286920547, + 0.006921254098415375, + 0.004416528157889843, + 0.011009964160621166, + 0.005005661863833666, + 0.005765127949416637, + 0.021874986588954926, + 0.01320636086165905, + 0.045913320034742355, + 0.016226552426815033, + 0.019272079691290855, + 0.05913256108760834, + 0.017630260437726974, + 0.024807967245578766, + 0.009084601886570454, + 0.010784994810819626, + 0.024901578202843666, + 0.004924408625811338, + 0.008287261240184307, + 0.01069561019539833, + 0.005207618232816458, + 0.01740710437297821, + 0.017714472487568855, + 0.01795411854982376, + 0.029060708358883858, + 0.015405354090034962, + 0.01395792979747057, + 0.02541426196694374, + 0.019624779000878334, + 0.023472033441066742, + 0.003118684282526374, + 0.003696920583024621, + 0.004594951402395964, + 0.0050937035121023655, + 0.0033359730150550604, + 0.0041744401678442955, + 0.004634421318769455, + 0.003160713938996196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.6327899694442749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022011594846844673 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.0022011594846844673, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022011594846844673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05954359099268913, + 0.17151853442192078, + 0.025229500606656075, + 0.12804749608039856, + 0.05694016069173813, + 0.08168640732765198, + 0.23138488829135895, + 0.1317102015018463, + 0.09003929793834686, + 0.02389988861978054 + ], + "credit_map": [ + 0.0052873180247843266, + 0.006364504341036081, + 0.012010178528726101, + 0.005374876782298088, + 0.004355395212769508, + 0.0064720516093075275, + 0.0046433075331151485, + 0.0150359608232975, + 0.011799877509474754, + 0.00822931807488203, + 0.05743049085140228, + 0.008309978991746902, + 0.006245801690965891, + 0.010473914444446564, + 0.005742599721997976, + 0.06328655779361725, + 0.004527502693235874, + 0.003299762262031436, + 0.0038068818394094706, + 0.003227237379178405, + 0.002284139161929488, + 0.0019894440192729235, + 0.0026945709250867367, + 0.003399962792173028, + 0.023403286933898926, + 0.005593080539256334, + 0.02709696628153324, + 0.008434356190264225, + 0.006958164740353823, + 0.02916558086872101, + 0.007905124686658382, + 0.01949094422161579, + 0.0055181169882416725, + 0.002841325243934989, + 0.019154135137796402, + 0.004401968326419592, + 0.0030827100854367018, + 0.014404736459255219, + 0.0039060257840901613, + 0.0036311401054263115, + 0.01596737839281559, + 0.004750980995595455, + 0.016824152320623398, + 0.006555554922670126, + 0.00535375764593482, + 0.016381828114390373, + 0.004620120394974947, + 0.011232640594244003, + 0.025477251037955284, + 0.017073893919587135, + 0.04609905555844307, + 0.01966828852891922, + 0.02057027630507946, + 0.0466063916683197, + 0.018364477902650833, + 0.03752525895833969, + 0.012539984658360481, + 0.009860713966190815, + 0.0375259667634964, + 0.00667713163420558, + 0.012972443364560604, + 0.032216738909482956, + 0.009410937316715717, + 0.010506289079785347, + 0.01009596511721611, + 0.007331119850277901, + 0.02199026197195053, + 0.008595682680606842, + 0.006676576565951109, + 0.011511532589793205, + 0.008723732084035873, + 0.015114428475499153, + 0.002088430803269148, + 0.0025698270183056593, + 0.0032150978222489357, + 0.004104457329958677, + 0.0022403665352612734, + 0.003412797814235091, + 0.003944477066397667, + 0.002324434695765376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.0022011594846844673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031980343628674746 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.0031980343628674746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031980343628674746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10486174374818802, + 0.06876225769519806, + 0.029615262523293495, + 0.1647159606218338, + 0.017575904726982117, + 0.17770132422447205, + 0.24347460269927979, + 0.12050438672304153, + 0.037135832011699677, + 0.035652726888656616 + ], + "credit_map": [ + 0.008469883352518082, + 0.006657570134848356, + 0.027312416583299637, + 0.007287102751433849, + 0.005245640873908997, + 0.011706712655723095, + 0.00659201480448246, + 0.03159039840102196, + 0.004731748718768358, + 0.004087309818714857, + 0.018860317766666412, + 0.005322832148522139, + 0.003575629089027643, + 0.005927993915975094, + 0.004551328253000975, + 0.021705102175474167, + 0.007141884882003069, + 0.0019760329741984606, + 0.005469759460538626, + 0.002275430364534259, + 0.0018450942588970065, + 0.0037316533271223307, + 0.0021101455204188824, + 0.005065261851996183, + 0.03195948898792267, + 0.005095364525914192, + 0.03372163698077202, + 0.008921022526919842, + 0.006966100540012121, + 0.0412580631673336, + 0.010982907377183437, + 0.025811370462179184, + 0.0017727825324982405, + 0.0010867089731618762, + 0.004085964057594538, + 0.001715311547741294, + 0.0010696977842599154, + 0.00430611427873373, + 0.0017442391254007816, + 0.001795086427591741, + 0.035242561250925064, + 0.011169236153364182, + 0.033871062099933624, + 0.01641496829688549, + 0.01004785019904375, + 0.035262465476989746, + 0.01040681917220354, + 0.025286374613642693, + 0.030763551592826843, + 0.014486776664853096, + 0.05148148536682129, + 0.017863724380731583, + 0.017982620745897293, + 0.05529610440135002, + 0.018241487443447113, + 0.0373588465154171, + 0.014891644939780235, + 0.017008857801556587, + 0.015457903034985065, + 0.009876717813313007, + 0.01809738017618656, + 0.017475971952080727, + 0.011447428725659847, + 0.016248485073447227, + 0.003266348736360669, + 0.0047319745644927025, + 0.007428121753036976, + 0.0035885281395167112, + 0.0040410347282886505, + 0.005960351321846247, + 0.004389935173094273, + 0.003729537595063448, + 0.004114734008908272, + 0.003512760391458869, + 0.004863508511334658, + 0.004776085261255503, + 0.003723799018189311, + 0.005631197709590197, + 0.005790557712316513, + 0.003240082645788789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0031980343628674746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031291097402573 + ], + "opposite_logprob_tokens": [ + -2.3503129482269287 + ], + "logprob_reward": -0.10031291097402573, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031291097402573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08410008996725082, + 0.27319833636283875, + 0.03709136322140694, + 0.045287635177373886, + 0.21027089655399323, + 0.09995431452989578, + 0.0635160431265831, + 0.07067395001649857, + 0.09609576314687729, + 0.019811667501926422 + ], + "credit_map": [ + 0.006526016164571047, + 0.005352293141186237, + 0.023135358467698097, + 0.005553173832595348, + 0.003917382098734379, + 0.007889729924499989, + 0.0052093141712248325, + 0.026516826823353767, + 0.017851129174232483, + 0.009861195459961891, + 0.09697120636701584, + 0.01110985316336155, + 0.006374292075634003, + 0.014848859049379826, + 0.007520781364291906, + 0.10866101086139679, + 0.0048273890279233456, + 0.0035388339310884476, + 0.006061865948140621, + 0.003369266400113702, + 0.003354446729645133, + 0.004692758433520794, + 0.0033060472924262285, + 0.007940755225718021, + 0.008042515255510807, + 0.003504761727526784, + 0.008372382260859013, + 0.003942482639104128, + 0.004177449736744165, + 0.007781691383570433, + 0.0036966102197766304, + 0.005769744049757719, + 0.015634704381227493, + 0.0034188309218734503, + 0.08361325412988663, + 0.008740559220314026, + 0.005802084691822529, + 0.0794423520565033, + 0.004377282690256834, + 0.009241815656423569, + 0.017342936247587204, + 0.008860190398991108, + 0.01633584126830101, + 0.011931415647268295, + 0.006979586556553841, + 0.01689229905605316, + 0.007465285714715719, + 0.014146754518151283, + 0.005412383005023003, + 0.005323655903339386, + 0.00751540157943964, + 0.007665806449949741, + 0.007780297659337521, + 0.013004058040678501, + 0.007196755614131689, + 0.009617685340344906, + 0.006972244940698147, + 0.011654243804514408, + 0.010123791173100471, + 0.005989153403788805, + 0.009298519231379032, + 0.010439850389957428, + 0.006051208823919296, + 0.010144935920834541, + 0.012411179021000862, + 0.01002774853259325, + 0.014425943605601788, + 0.010778414085507393, + 0.007958226837217808, + 0.01671220175921917, + 0.009487953037023544, + 0.014294096268713474, + 0.001850337372161448, + 0.0022595070768147707, + 0.002972703892737627, + 0.0031519089825451374, + 0.0021048756316304207, + 0.0031916419975459576, + 0.0022034950088709593, + 0.0020771967247128487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.10031291097402573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, downward close-up follows an ancient, yellowed human skull as it slowly but steadily descends to the bottom of a dark, murky depth. A faint, almost invisible beam of light penetrates from the distant surface, barely framing the skull's bony surface, punctuated by microscopic cracks and the erosion of time. As it sinks, a dense, continuous stream of tiny bubbles emanates from its deep, hollow eye sockets. As they rise, in stark contrast to the skull's downward trajectory, these bubbles capture the glimmer of light, briefly flashing a ghostly silver glow before dissipating into the surrounding darkness. The entire image, shrouded in a profoundly cold hue, creates a dead, mysterious, and unsettling atmosphere, as if silently telling a forgotten tale of eternal perdition.", + "seed": 15282, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15282_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027004318637773396, + "vlm_logprob": -0.027004318637773396, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.027004318637773396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01573292538523674 + ], + "opposite_logprob_tokens": [ + -4.390732765197754 + ], + "logprob_reward": -0.01573292538523674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01573292538523674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17268550395965576, + 0.13840462267398834, + 0.08176415413618088, + 0.09044533222913742, + 0.18824703991413116, + 0.09697123616933823, + 0.0808635801076889, + 0.06140313297510147, + 0.04324069619178772, + 0.04597463086247444 + ], + "credit_map": [ + 0.016270741820335388, + 0.01525899488478899, + 0.022956890985369682, + 0.01808253489434719, + 0.018798159435391426, + 0.023134766146540642, + 0.02787575125694275, + 0.03030768223106861, + 0.009600038640201092, + 0.018225762993097305, + 0.019430551677942276, + 0.013968370854854584, + 0.017202727496623993, + 0.016257524490356445, + 0.02034299261868, + 0.023376675322651863, + 0.0093693183735013, + 0.00954076461493969, + 0.008660449646413326, + 0.014042901806533337, + 0.013124403543770313, + 0.004872530233114958, + 0.013004286214709282, + 0.009149497374892235, + 0.010585043579339981, + 0.008572116494178772, + 0.01752612553536892, + 0.013262641616165638, + 0.00970620010048151, + 0.007977000437676907, + 0.009369716979563236, + 0.013446500524878502, + 0.02074122428894043, + 0.018823638558387756, + 0.032724879682064056, + 0.019388223066926003, + 0.009412744082510471, + 0.014213540591299534, + 0.03095669113099575, + 0.04198611527681351, + 0.009041349403560162, + 0.007961050607264042, + 0.01592152751982212, + 0.014001724310219288, + 0.007551400922238827, + 0.012325769290328026, + 0.017822792753577232, + 0.012345633469522, + 0.010200593620538712, + 0.010797728784382343, + 0.010687451809644699, + 0.00825987383723259, + 0.009222351014614105, + 0.007459980435669422, + 0.0118775749579072, + 0.01235803123563528, + 0.005771196447312832, + 0.01171000674366951, + 0.010075876489281654, + 0.005658529698848724, + 0.007188436575233936, + 0.008859632536768913, + 0.005325636360794306, + 0.006813823711127043, + 0.0039285034872591496, + 0.008654575794935226, + 0.0053985025733709335, + 0.004594098310917616, + 0.006156753282994032, + 0.004291878081858158, + 0.005613775923848152, + 0.004602612461894751, + 0.0037413127720355988, + 0.0031308168545365334, + 0.006552622653543949, + 0.002628917107358575, + 0.003818535478785634, + 0.011658240109682083, + 0.005703344475477934, + 0.008740842342376709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01573292538523674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899461686611176 + ], + "opposite_logprob_tokens": [ + -2.5789945125579834 + ], + "logprob_reward": -0.07899461686611176, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899461686611176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14604514837265015, + 0.06826519221067429, + 0.05489572882652283, + 0.12220517545938492, + 0.14730700850486755, + 0.15596793591976166, + 0.11977852135896683, + 0.05855962634086609, + 0.054561465978622437, + 0.07241416722536087 + ], + "credit_map": [ + 0.02389231137931347, + 0.012612392194569111, + 0.017174018546938896, + 0.013062438927590847, + 0.008979199454188347, + 0.025672517716884613, + 0.01290646567940712, + 0.03174581751227379, + 0.008967899717390537, + 0.006633535493165255, + 0.014181760139763355, + 0.006497356574982405, + 0.006087175104767084, + 0.007873929105699062, + 0.007830925285816193, + 0.010192605666816235, + 0.005321920849382877, + 0.004857460502535105, + 0.009746630676090717, + 0.004519715439528227, + 0.011127220466732979, + 0.003230733098462224, + 0.0065195211209356785, + 0.0095725292339921, + 0.017130572348833084, + 0.007038596551865339, + 0.027832861989736557, + 0.012688841670751572, + 0.010332324542105198, + 0.010050764307379723, + 0.008694084361195564, + 0.02843713015317917, + 0.013594607822597027, + 0.020066257566213608, + 0.018827350810170174, + 0.018751800060272217, + 0.010637043043971062, + 0.011412901803851128, + 0.033698469400405884, + 0.020318569615483284, + 0.01360981073230505, + 0.012895067222416401, + 0.01894981600344181, + 0.019207989796996117, + 0.016996052116155624, + 0.029584646224975586, + 0.026397082954645157, + 0.018327470868825912, + 0.009747492149472237, + 0.018725011497735977, + 0.014193258248269558, + 0.012126724235713482, + 0.020128361880779266, + 0.01686949096620083, + 0.012320424430072308, + 0.015667753294110298, + 0.006007352843880653, + 0.008705370128154755, + 0.010432601906359196, + 0.005056341644376516, + 0.007497643120586872, + 0.007891581393778324, + 0.0061879889108240604, + 0.006780743598937988, + 0.00490258214995265, + 0.007241274230182171, + 0.007668875157833099, + 0.005480783525854349, + 0.008895748294889927, + 0.00774381123483181, + 0.007093841675668955, + 0.005534547846764326, + 0.006033746991306543, + 0.004993520211428404, + 0.011310468427836895, + 0.0037580288480967283, + 0.006178292445838451, + 0.020139217376708984, + 0.009030678309500217, + 0.010970210656523705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.07899461686611176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024944401811808348 + ], + "opposite_logprob_tokens": [ + -6.0024943351745605 + ], + "logprob_reward": -0.0024944401811808348, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024944401811808348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08893750607967377, + 0.11753752082586288, + 0.05493711307644844, + 0.06994476914405823, + 0.11166031658649445, + 0.12072425335645676, + 0.1501389443874359, + 0.08414147049188614, + 0.11622290313243866, + 0.08575522154569626 + ], + "credit_map": [ + 0.017785467207431793, + 0.008573394268751144, + 0.013386487029492855, + 0.008698566816747189, + 0.009862647391855717, + 0.007876682095229626, + 0.013367344625294209, + 0.009386912919580936, + 0.00818379782140255, + 0.016945896670222282, + 0.01564006693661213, + 0.011802298948168755, + 0.016520053148269653, + 0.014274891465902328, + 0.016339361667633057, + 0.017831141129136086, + 0.005342388525605202, + 0.006063934415578842, + 0.01202895026654005, + 0.004959999583661556, + 0.007790714502334595, + 0.00319627090357244, + 0.0045667486265301704, + 0.010988103225827217, + 0.009893646463751793, + 0.007309316657483578, + 0.012643720023334026, + 0.009798920713365078, + 0.0060820733197033405, + 0.007959235459566116, + 0.006238069850951433, + 0.010019785724580288, + 0.010826682671904564, + 0.017757730558514595, + 0.011807396076619625, + 0.016599971801042557, + 0.011377543210983276, + 0.009456796571612358, + 0.02391894720494747, + 0.009915251284837723, + 0.010226793587207794, + 0.011060199700295925, + 0.014444414526224136, + 0.01717357523739338, + 0.011821080930531025, + 0.019907381385564804, + 0.020812107250094414, + 0.015278696082532406, + 0.017495382577180862, + 0.020937489345669746, + 0.0156609695404768, + 0.014349319040775299, + 0.027644408866763115, + 0.019095012918114662, + 0.01814895309507847, + 0.01680741272866726, + 0.005733800120651722, + 0.011258122511208057, + 0.017222529277205467, + 0.009571567177772522, + 0.010591993108391762, + 0.014585517346858978, + 0.008768068626523018, + 0.006409873254597187, + 0.011190083809196949, + 0.021307701244950294, + 0.017927363514900208, + 0.011508936993777752, + 0.014488042332231998, + 0.013608909212052822, + 0.01435527391731739, + 0.011836593970656395, + 0.007368430960923433, + 0.006680791731923819, + 0.013246678747236729, + 0.0060411496087908745, + 0.01003915723413229, + 0.020173363387584686, + 0.011598797515034676, + 0.010606854222714901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.0024944401811808348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0619589127600193, + 0.01949952356517315, + 0.12099718302488327, + 0.07318613678216934, + 0.12580060958862305, + 0.1522444635629654, + 0.05257048457860947, + 0.13295921683311462, + 0.15613342821598053, + 0.10465003550052643 + ], + "credit_map": [ + 0.009152918122708797, + 0.004607832059264183, + 0.008754358626902103, + 0.005566212814301252, + 0.0044184583239257336, + 0.010239948518574238, + 0.007569120731204748, + 0.011650062166154385, + 0.0021836746018379927, + 0.0022098023910075426, + 0.00239731278270483, + 0.0023095787037163973, + 0.002367405453696847, + 0.0020576852839440107, + 0.002988050226122141, + 0.002986014122143388, + 0.012582984752953053, + 0.005250811576843262, + 0.030057864263653755, + 0.017213068902492523, + 0.006261101923882961, + 0.006287447642534971, + 0.01014239713549614, + 0.03320150077342987, + 0.008226404897868633, + 0.008497080765664577, + 0.013732152990996838, + 0.006948645226657391, + 0.00812453031539917, + 0.00659765163436532, + 0.006443028338253498, + 0.014616641215980053, + 0.013432677835226059, + 0.019178535789251328, + 0.015240977518260479, + 0.014693926088511944, + 0.012179629877209663, + 0.014078590087592602, + 0.021138852462172508, + 0.015857430174946785, + 0.016856931149959564, + 0.012671610340476036, + 0.020538948476314545, + 0.01875229738652706, + 0.013985564932227135, + 0.027651114389300346, + 0.02031649835407734, + 0.021471502259373665, + 0.004669892136007547, + 0.009158430621027946, + 0.007414365652948618, + 0.0061742207035422325, + 0.008947874419391155, + 0.0054559786804020405, + 0.006661897990852594, + 0.004087825771421194, + 0.0127614950761199, + 0.015329834073781967, + 0.016891449689865112, + 0.02144758775830269, + 0.01512706559151411, + 0.021114660426974297, + 0.01631179265677929, + 0.013975338079035282, + 0.014032057486474514, + 0.02892422303557396, + 0.02361837588250637, + 0.016121482476592064, + 0.020172951743006706, + 0.018818065524101257, + 0.017730658873915672, + 0.016715604811906815, + 0.009023803286254406, + 0.006277826149016619, + 0.021692808717489243, + 0.006135352421551943, + 0.009245582856237888, + 0.02610555849969387, + 0.01193106546998024, + 0.014238039962947369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.004100012592971325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369959816336632 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.03369959816336632, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369959816336632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10533910989761353, + 0.029216662049293518, + 0.06135909631848335, + 0.11929575353860855, + 0.06393422931432724, + 0.18883399665355682, + 0.08275733143091202, + 0.14264270663261414, + 0.15714919567108154, + 0.04947197809815407 + ], + "credit_map": [ + 0.013475850224494934, + 0.009066547267138958, + 0.011474311351776123, + 0.010005911812186241, + 0.007796561811119318, + 0.02009483613073826, + 0.010602887719869614, + 0.022822199389338493, + 0.003191548865288496, + 0.0035514323972165585, + 0.0037348996847867966, + 0.003661396214738488, + 0.0033387313596904278, + 0.0024525162298232317, + 0.004923934582620859, + 0.004362198989838362, + 0.007598580326884985, + 0.006466934457421303, + 0.007214620243757963, + 0.009685203433036804, + 0.01009021420031786, + 0.003899795701727271, + 0.008536032401025295, + 0.007867714390158653, + 0.014579642564058304, + 0.010269026271998882, + 0.02247639372944832, + 0.012373100966215134, + 0.011797748506069183, + 0.012877493165433407, + 0.010663625784218311, + 0.02425871230661869, + 0.007075724191963673, + 0.005919148679822683, + 0.00906887836754322, + 0.008608438074588776, + 0.0044572800397872925, + 0.005733024794608355, + 0.011249619536101818, + 0.011822113767266273, + 0.019823230803012848, + 0.015222197398543358, + 0.024697141721844673, + 0.017743822187185287, + 0.019623931497335434, + 0.042605094611644745, + 0.02278958447277546, + 0.02632897160947323, + 0.010456804186105728, + 0.011448358185589314, + 0.011959776282310486, + 0.010165175423026085, + 0.011011200957000256, + 0.007885947823524475, + 0.010196168906986713, + 0.009633888490498066, + 0.014111694879829884, + 0.016498306766152382, + 0.018388310447335243, + 0.022753972560167313, + 0.01490194071084261, + 0.022787097841501236, + 0.017481070011854172, + 0.015720300376415253, + 0.013802378438413143, + 0.02909054048359394, + 0.02226845920085907, + 0.016763649880886078, + 0.022091930732131004, + 0.017513329163193703, + 0.019175397232174873, + 0.01644349843263626, + 0.004262970294803381, + 0.0030917366966605186, + 0.00856856070458889, + 0.002445058198645711, + 0.00429444620385766, + 0.013774828985333443, + 0.006120327394455671, + 0.006914048921316862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.03369959816336632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, downward close-up follows an ancient, yellowed human skull as it slowly but steadily descends to the bottom of a dark, murky depth. A faint, almost invisible beam of light penetrates from the distant surface, barely framing the skull's bony surface, punctuated by microscopic cracks and the erosion of time. As it sinks, a dense, continuous stream of tiny bubbles emanates from its deep, hollow eye sockets. As they rise, in stark contrast to the skull's downward trajectory, these bubbles capture the glimmer of light, briefly flashing a ghostly silver glow before dissipating into the surrounding darkness. The entire image, shrouded in a profoundly cold hue, creates a dead, mysterious, and unsettling atmosphere, as if silently telling a forgotten tale of eternal perdition.", + "seed": 15283, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_13/index_2741_seed_15283_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03076570192351937, + "vlm_logprob": -0.03076570192351937, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03076570192351937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009147756733000278 + ], + "opposite_logprob_tokens": [ + -5.134147644042969 + ], + "logprob_reward": -0.009147756733000278, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009147756733000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20683056116104126, + 0.1094299778342247, + 0.10593720525503159, + 0.11721019446849823, + 0.09866026043891907, + 0.05180571228265762, + 0.09897482395172119, + 0.06421568989753723, + 0.055195339024066925, + 0.09174030274152756 + ], + "credit_map": [ + 0.04057741537690163, + 0.03445809707045555, + 0.03287042677402496, + 0.03134055808186531, + 0.015490835532546043, + 0.012949418276548386, + 0.021248944103717804, + 0.01789485290646553, + 0.012944893911480904, + 0.014876405708491802, + 0.014001722447574139, + 0.014996057376265526, + 0.016427839174866676, + 0.009052322246134281, + 0.018480904400348663, + 0.008649826049804688, + 0.014395739883184433, + 0.011311125010251999, + 0.014289427548646927, + 0.0148379011079669, + 0.014737062156200409, + 0.007558893412351608, + 0.020534353330731392, + 0.008272698149085045, + 0.014400666579604149, + 0.018796870484948158, + 0.017141133546829224, + 0.019933732226490974, + 0.01222730427980423, + 0.009132915176451206, + 0.014840962365269661, + 0.010736598633229733, + 0.010101811960339546, + 0.012342859990894794, + 0.011060296557843685, + 0.0139627018943429, + 0.014382602646946907, + 0.009987548924982548, + 0.016864193603396416, + 0.009958240203559399, + 0.007081884425133467, + 0.005100687500089407, + 0.00810635183006525, + 0.007997238077223301, + 0.004269278142601252, + 0.00369689567014575, + 0.008817262016236782, + 0.006736110895872116, + 0.00984848290681839, + 0.01632779836654663, + 0.011605307459831238, + 0.012290647253394127, + 0.017495576292276382, + 0.014172409661114216, + 0.009731782600283623, + 0.007502813823521137, + 0.009323877282440662, + 0.005963375326246023, + 0.00977107509970665, + 0.009817912243306637, + 0.00595645559951663, + 0.0054642255418002605, + 0.006975527387112379, + 0.01094323955476284, + 0.006177330855280161, + 0.011116954497992992, + 0.007149951998144388, + 0.007881337776780128, + 0.006396181881427765, + 0.005028425250202417, + 0.005792739801108837, + 0.005652417428791523, + 0.012813808396458626, + 0.008496289141476154, + 0.010728996247053146, + 0.014135892502963543, + 0.013392852619290352, + 0.010200128890573978, + 0.011930804699659348, + 0.010041522793471813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009147756733000278 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294075846672058 + ], + "opposite_logprob_tokens": [ + -2.237940788269043 + ], + "logprob_reward": -0.11294075846672058, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294075846672058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09130831807851791, + 0.07086753845214844, + 0.07177098095417023, + 0.05968211963772774, + 0.08606521785259247, + 0.073492132127285, + 0.06749337911605835, + 0.26158925890922546, + 0.14674288034439087, + 0.07098817080259323 + ], + "credit_map": [ + 0.013989226892590523, + 0.01943971961736679, + 0.014598657377064228, + 0.01699928008019924, + 0.007291149813681841, + 0.0054233078844845295, + 0.008415509015321732, + 0.005151471588760614, + 0.016088832169771194, + 0.008475909940898418, + 0.012518213130533695, + 0.009677101857960224, + 0.005870748311281204, + 0.005209252703934908, + 0.007328064180910587, + 0.005699418019503355, + 0.009768637828528881, + 0.009621899574995041, + 0.010330219753086567, + 0.013569330796599388, + 0.006049139425158501, + 0.005409093573689461, + 0.009425856173038483, + 0.00759680662304163, + 0.007740855682641268, + 0.010255293920636177, + 0.007785839028656483, + 0.007606253959238529, + 0.006178995128720999, + 0.003709036624059081, + 0.010128860361874104, + 0.006276980042457581, + 0.01084916666150093, + 0.012478317134082317, + 0.011381794698536396, + 0.014476792886853218, + 0.005613934248685837, + 0.005260174628347158, + 0.014799054712057114, + 0.011205983348190784, + 0.011326968669891357, + 0.008149396628141403, + 0.008195740170776844, + 0.012208727188408375, + 0.008024911396205425, + 0.00764109892770648, + 0.012034772895276546, + 0.00591051671653986, + 0.006004496477544308, + 0.008526276797056198, + 0.008124440908432007, + 0.00906654167920351, + 0.015778042376041412, + 0.0071360100992023945, + 0.007646172307431698, + 0.005211393348872662, + 0.0417056530714035, + 0.03490520641207695, + 0.03899753466248512, + 0.04844256117939949, + 0.01875772513449192, + 0.016810862347483635, + 0.02956736460328102, + 0.032402366399765015, + 0.015050763264298439, + 0.02287840098142624, + 0.02416827529668808, + 0.020743008702993393, + 0.010682471096515656, + 0.013662988319993019, + 0.023027770221233368, + 0.01652919501066208, + 0.005951752886176109, + 0.008488088846206665, + 0.009255520068109035, + 0.013686076737940311, + 0.007717817556113005, + 0.006629333831369877, + 0.010007344186306, + 0.009252231568098068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.11294075846672058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09285757690668106, + 0.04873954877257347, + 0.10851630568504333, + 0.08371896296739578, + 0.09973476827144623, + 0.07884413748979568, + 0.13925258815288544, + 0.1051076278090477, + 0.18175071477890015, + 0.06147780269384384 + ], + "credit_map": [ + 0.01656702347099781, + 0.016978221014142036, + 0.013755260035395622, + 0.01781608909368515, + 0.005698713473975658, + 0.005816350225359201, + 0.00956348143517971, + 0.0066624293103814125, + 0.010213792324066162, + 0.004275115206837654, + 0.0068187471479177475, + 0.005308124236762524, + 0.003483052598312497, + 0.003548248438164592, + 0.007890874519944191, + 0.007201595231890678, + 0.016397764906287193, + 0.022939030081033707, + 0.01643381454050541, + 0.019077682867646217, + 0.007594721391797066, + 0.006370075047016144, + 0.009753795340657234, + 0.009949416853487492, + 0.011203676462173462, + 0.016226235777139664, + 0.011314944364130497, + 0.01308455504477024, + 0.008617989718914032, + 0.005702686030417681, + 0.012530535459518433, + 0.005038340110331774, + 0.01194752100855112, + 0.010842622257769108, + 0.013219618238508701, + 0.013924723491072655, + 0.01272713765501976, + 0.009268874302506447, + 0.018604613840579987, + 0.009199646301567554, + 0.008415879681706429, + 0.009648560546338558, + 0.010761154815554619, + 0.012086128816008568, + 0.010626479983329773, + 0.008310018107295036, + 0.011792941950261593, + 0.007202960085123777, + 0.01247870922088623, + 0.022245382890105247, + 0.017362261191010475, + 0.014629796147346497, + 0.03182241693139076, + 0.019777581095695496, + 0.012378130108118057, + 0.008558299392461777, + 0.014421969652175903, + 0.016928035765886307, + 0.01546495221555233, + 0.017439067363739014, + 0.008823396638035774, + 0.008175485767424107, + 0.01239837147295475, + 0.011456331238150597, + 0.018109958618879318, + 0.027838433161377907, + 0.032037436962127686, + 0.023685721680521965, + 0.015739919617772102, + 0.01561736036092043, + 0.030592378228902817, + 0.018129494041204453, + 0.009260180406272411, + 0.007517229299992323, + 0.007798542268574238, + 0.00893637165427208, + 0.006975861266255379, + 0.005675062537193298, + 0.0047458503395318985, + 0.010568701662123203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.00525581045076251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14864188432693481, + 0.04206470400094986, + 0.08619426190853119, + 0.21531514823436737, + 0.09517347812652588, + 0.06998052448034286, + 0.12726272642612457, + 0.08017661422491074, + 0.05377279594540596, + 0.08141782134771347 + ], + "credit_map": [ + 0.023476844653487206, + 0.02999977208673954, + 0.022899053990840912, + 0.022958775982260704, + 0.010834148153662682, + 0.01229209452867508, + 0.015047836117446423, + 0.011133360676467419, + 0.006360047962516546, + 0.006168138701468706, + 0.00628711236640811, + 0.006054258439689875, + 0.003943983465433121, + 0.003338281996548176, + 0.005366788245737553, + 0.004546096082776785, + 0.012420454993844032, + 0.018224000930786133, + 0.012873795814812183, + 0.013578151352703571, + 0.005263738799840212, + 0.004448868799954653, + 0.013482085429131985, + 0.00590317090973258, + 0.03126171976327896, + 0.047006480395793915, + 0.030482379719614983, + 0.036255717277526855, + 0.019341809675097466, + 0.012255086563527584, + 0.02217458002269268, + 0.01653738133609295, + 0.015176719054579735, + 0.014718501828610897, + 0.013443929143249989, + 0.012208441272377968, + 0.008184618316590786, + 0.007197155617177486, + 0.014303743839263916, + 0.009940382093191147, + 0.008556882850825787, + 0.008763189427554607, + 0.011333293281495571, + 0.015736671164631844, + 0.005420403089374304, + 0.005090455990284681, + 0.006632058881223202, + 0.008447574451565742, + 0.01071701105684042, + 0.021564293652772903, + 0.013848181813955307, + 0.016152188181877136, + 0.0249264407902956, + 0.01728047803044319, + 0.013292098417878151, + 0.009482042863965034, + 0.011929566971957684, + 0.009122888557612896, + 0.011636168695986271, + 0.014227804727852345, + 0.007381765637546778, + 0.007533966563642025, + 0.009372230619192123, + 0.008972221985459328, + 0.006023114547133446, + 0.01224483735859394, + 0.008021799847483635, + 0.00670795701444149, + 0.00561289768666029, + 0.004863614682108164, + 0.005203046835958958, + 0.005095528904348612, + 0.010625960305333138, + 0.011192663572728634, + 0.009127759374678135, + 0.014570744708180428, + 0.00836618896573782, + 0.007957731373608112, + 0.010654154233634472, + 0.008922629058361053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient, yellowed human skull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skull moving downward to the bottom of a dark, murky depth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny bubbles emanate from the skull's deep, hollow eye sockets and rise upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a profoundly cold hue with a faint beam of light from the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a tilted, downward close-up following the skull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289239034056664 + ], + "opposite_logprob_tokens": [ + -3.773289203643799 + ], + "logprob_reward": -0.023289239034056664, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289239034056664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1058688536286354, + 0.06256117671728134, + 0.13551227748394012, + 0.17857936024665833, + 0.062379322946071625, + 0.10329210758209229, + 0.10653485357761383, + 0.07321333140134811, + 0.08748798072338104, + 0.08457068353891373 + ], + "credit_map": [ + 0.015591769479215145, + 0.021136920899152756, + 0.01619143597781658, + 0.020762277767062187, + 0.008544634096324444, + 0.0064481343142688274, + 0.00991640705615282, + 0.007277274038642645, + 0.0158169437199831, + 0.008172689937055111, + 0.010578456334769726, + 0.007363183423876762, + 0.004255285486578941, + 0.0029592004138976336, + 0.007836790755391121, + 0.005578625947237015, + 0.021492179483175278, + 0.01646135002374649, + 0.020416732877492905, + 0.02378682978451252, + 0.012169744819402695, + 0.009413379244506359, + 0.016645662486553192, + 0.015126395970582962, + 0.025846635922789574, + 0.03454545512795448, + 0.02369164302945137, + 0.033858414739370346, + 0.019769709557294846, + 0.009748440235853195, + 0.021945895627141, + 0.009173167869448662, + 0.007687210571020842, + 0.008279790170490742, + 0.007720841560512781, + 0.008863919414579868, + 0.005937913432717323, + 0.005595400929450989, + 0.010696373879909515, + 0.007597870193421841, + 0.01228944119066, + 0.013892951421439648, + 0.015239054337143898, + 0.019110124558210373, + 0.011458114720880985, + 0.008939174003899097, + 0.012180118821561337, + 0.010183120146393776, + 0.012608127668499947, + 0.015768494457006454, + 0.014206554740667343, + 0.013734788633883, + 0.015559581108391285, + 0.013222607783973217, + 0.011199637316167355, + 0.010235062800347805, + 0.0067948163487017155, + 0.011438287794589996, + 0.011541160754859447, + 0.010998492129147053, + 0.007795573677867651, + 0.008911567740142345, + 0.009639734402298927, + 0.006093700882047415, + 0.009677860885858536, + 0.009292248636484146, + 0.011631298810243607, + 0.01252845861017704, + 0.00975379254668951, + 0.011279440484941006, + 0.011546351946890354, + 0.01177852414548397, + 0.00972835160791874, + 0.01467104721814394, + 0.00861078780144453, + 0.013680468313395977, + 0.011853599920868874, + 0.006588593125343323, + 0.006080579478293657, + 0.013357254676520824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.023289239034056664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_tensor([3636, 3636])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_tensor([3636, 3636])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..25a932f2103fd35bda8dd810141ec4425c31243d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_tensor([3636, 3636])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static wide-angle shot, shot from a level angle, captures the vast expanse of the ocean amidst a torrential downpour. A small wooden fishing boat, painted a faded red, stands as the sole focal point. Beneath a sky shrouded in thick, leaden clouds, a dense curtain of rain cascades down from the sky, completely blurring the line between sea and sky. The dark gray waters churn, whipping up surging waves capped with white foam. The boat appears vulnerable to this violent force. A massive wave lifts the bow almost perpendicular to the surface, only to slam it back into the trough, sending up a shower of white spray. The boat desperately sways from side to side, struggling to maintain its balance. The entire scene, bathed in cool, even, diffuse light, appears remarkably clean and crisp. The high shutter speed freezes every moment, making every splash and tumbling wave visible, devoid of motion blur. This extremely clear visual style contrasts sharply with the violent turbulence of the scene, highlighting the ruthless power of nature and the loneliness and tenacity of life struggling within it, creating a grand, cold and visually impactful atmosphere.", + "seed": 12716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_3636_seed_12716_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.472504872456193, + "vlm_logprob": -0.472504872456193, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.472504872456193, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11178011447191238 + ], + "opposite_logprob_tokens": [ + -2.7367801666259766 + ], + "logprob_reward": -0.11178011447191238, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11178011447191238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20261535048484802, + 0.03881431370973587, + 0.03813421353697777, + 0.04221263900399208, + 0.09051469713449478, + 0.0752175971865654, + 0.08419213443994522, + 0.06592531502246857, + 0.2742137312889099, + 0.08815999329090118 + ], + "credit_map": [ + 0.015665769577026367, + 0.03469580411911011, + 0.024639097973704338, + 0.04180155694484711, + 0.031090624630451202, + 0.022499458864331245, + 0.01498990785330534, + 0.01723312772810459, + 0.0035399922635406256, + 0.004949749913066626, + 0.006746266968548298, + 0.0050521548837423325, + 0.0045495410449802876, + 0.005152769852429628, + 0.005077641922980547, + 0.003746198723092675, + 0.0035009789280593395, + 0.006462656892836094, + 0.004182691220194101, + 0.005379406735301018, + 0.008555339649319649, + 0.004115895368158817, + 0.0030819035600870848, + 0.0028553421143442392, + 0.004034674726426601, + 0.005789837334305048, + 0.007413178216665983, + 0.004508881364017725, + 0.0052877068519592285, + 0.007431552279740572, + 0.004798391368240118, + 0.0029484175611287355, + 0.007435755804181099, + 0.014942869544029236, + 0.014300517737865448, + 0.018092747777700424, + 0.014203416183590889, + 0.009432755410671234, + 0.005763539578765631, + 0.006343090906739235, + 0.007365988101810217, + 0.010382020846009254, + 0.014607134275138378, + 0.016041157767176628, + 0.007841793820261955, + 0.00701091019436717, + 0.00651481281965971, + 0.005453781690448523, + 0.008185169659554958, + 0.017814790830016136, + 0.011465179733932018, + 0.0068717109970748425, + 0.01168578676879406, + 0.01235252246260643, + 0.008567760698497295, + 0.007249206770211458, + 0.0074956389144063, + 0.009940309450030327, + 0.008353549055755138, + 0.01110259722918272, + 0.009445552714169025, + 0.005902435164898634, + 0.007381092756986618, + 0.0063041457906365395, + 0.02426682598888874, + 0.03431358188390732, + 0.03294147178530693, + 0.04135613888502121, + 0.03877263888716698, + 0.032369863241910934, + 0.047147445380687714, + 0.023045778274536133, + 0.006474293302744627, + 0.011058943346142769, + 0.009074674919247627, + 0.015348813496530056, + 0.015760691836476326, + 0.011454356834292412, + 0.011282616294920444, + 0.007705611642450094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.11178011447191238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255536638200283 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.01255536638200283, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255536638200283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13216586410999298, + 0.1221599280834198, + 0.07500477135181427, + 0.08546233177185059, + 0.10963930934667587, + 0.0868358463048935, + 0.11375319957733154, + 0.10928969085216522, + 0.13545486330986023, + 0.0302341990172863 + ], + "credit_map": [ + 0.012270985171198845, + 0.02130930684506893, + 0.02044777199625969, + 0.017426880076527596, + 0.015538358129560947, + 0.015692828223109245, + 0.014652528800070286, + 0.014827199280261993, + 0.010121459141373634, + 0.014517880976200104, + 0.013211533427238464, + 0.019026849418878555, + 0.015848001465201378, + 0.014941840432584286, + 0.021603945642709732, + 0.012888416647911072, + 0.005550774279981852, + 0.010952435433864594, + 0.007318140473216772, + 0.008324210532009602, + 0.013896984979510307, + 0.009722288697957993, + 0.008964795619249344, + 0.010275139473378658, + 0.007525813765823841, + 0.010482284240424633, + 0.015488157980144024, + 0.010811200365424156, + 0.012312700040638447, + 0.00869032833725214, + 0.01191549003124237, + 0.008236357942223549, + 0.011393280699849129, + 0.014454741962254047, + 0.011875713244080544, + 0.01876807026565075, + 0.011910023167729378, + 0.010594329796731472, + 0.017646603286266327, + 0.012996548786759377, + 0.0076364316046237946, + 0.01465247105807066, + 0.012223823927342892, + 0.01548553816974163, + 0.010602652095258236, + 0.011567465029656887, + 0.007450887002050877, + 0.007216580677777529, + 0.008688714355230331, + 0.011941837146878242, + 0.015309450216591358, + 0.012390556745231152, + 0.009031662717461586, + 0.008915129117667675, + 0.03254224359989166, + 0.01493360847234726, + 0.010940306819975376, + 0.012762407772243023, + 0.020825721323490143, + 0.015130255371332169, + 0.013788973912596703, + 0.012204705737531185, + 0.01354055106639862, + 0.010096766985952854, + 0.01187349297106266, + 0.016189808025956154, + 0.018318723887205124, + 0.02243191935122013, + 0.01376827247440815, + 0.01182330772280693, + 0.028422174975275993, + 0.01262715458869934, + 0.0024406593292951584, + 0.003813820192590356, + 0.003587401704862714, + 0.005329974461346865, + 0.004163957666605711, + 0.002490411978214979, + 0.005034231580793858, + 0.0033737437333911657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01255536638200283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353973895311356 + ], + "opposite_logprob_tokens": [ + -3.6513540744781494 + ], + "logprob_reward": -0.026353973895311356, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353973895311356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18120622634887695, + 0.07478965818881989, + 0.042307984083890915, + 0.06511008739471436, + 0.2356528490781784, + 0.1873239427804947, + 0.07704564929008484, + 0.04149378836154938, + 0.06357593834400177, + 0.03149387240409851 + ], + "credit_map": [ + 0.0156544279307127, + 0.031431443989276886, + 0.021053172647953033, + 0.03709844499826431, + 0.02681981772184372, + 0.017999451607465744, + 0.014021007344126701, + 0.017128445208072662, + 0.0055550262331962585, + 0.010084780864417553, + 0.010943058878183365, + 0.007099359296262264, + 0.015274482779204845, + 0.010540123097598553, + 0.009258393198251724, + 0.006034435238689184, + 0.004350787959992886, + 0.0061470866203308105, + 0.006838821806013584, + 0.0044701071456074715, + 0.005488762632012367, + 0.004148462787270546, + 0.005829593166708946, + 0.005034363828599453, + 0.005327749997377396, + 0.010159310884773731, + 0.006378518417477608, + 0.008014693856239319, + 0.016085216775536537, + 0.008215353824198246, + 0.007029037456959486, + 0.003900208044797182, + 0.017417894676327705, + 0.028780100867152214, + 0.034468844532966614, + 0.04999057576060295, + 0.02478480339050293, + 0.028406459838151932, + 0.027869004756212234, + 0.023935167118906975, + 0.015701506286859512, + 0.026314619928598404, + 0.02828803099691868, + 0.031385473906993866, + 0.027061907574534416, + 0.023604437708854675, + 0.01735376939177513, + 0.01761418953537941, + 0.00705506419762969, + 0.013086099177598953, + 0.012560000643134117, + 0.01231525931507349, + 0.008413225412368774, + 0.007348451763391495, + 0.009037735871970654, + 0.007229808252304792, + 0.003985387273132801, + 0.00763769680634141, + 0.00598669471219182, + 0.004165873862802982, + 0.006877434905618429, + 0.004451496992260218, + 0.005421427544206381, + 0.0029677751008421183, + 0.005567676853388548, + 0.007802152540534735, + 0.008917250670492649, + 0.01032886654138565, + 0.008394930511713028, + 0.007383640855550766, + 0.009161879308521748, + 0.006019537337124348, + 0.0024048846680670977, + 0.004844409879297018, + 0.005210452247411013, + 0.005345024634152651, + 0.0042960732243955135, + 0.00250421860255301, + 0.003359320806339383, + 0.003529484150931239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.026353973895311356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983492851257324 + ], + "opposite_logprob_tokens": [ + -0.5233492851257324 + ], + "logprob_reward": -0.8983492851257324, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983492851257324, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13341370224952698, + 0.13251417875289917, + 0.14232148230075836, + 0.04244718328118324, + 0.19218501448631287, + 0.09873572736978531, + 0.06450208276510239, + 0.07516007870435715, + 0.0662812665104866, + 0.052439287304878235 + ], + "credit_map": [ + 0.01088322140276432, + 0.02262750267982483, + 0.022600233554840088, + 0.01769208163022995, + 0.02304721623659134, + 0.01753455027937889, + 0.01047146413475275, + 0.008557425811886787, + 0.011108232662081718, + 0.020291663706302643, + 0.011762809008359909, + 0.013646239414811134, + 0.036473896354436874, + 0.01847231388092041, + 0.010941719636321068, + 0.009817290119826794, + 0.011119275353848934, + 0.02065901830792427, + 0.013427969999611378, + 0.025927482172846794, + 0.028209703043103218, + 0.013294011354446411, + 0.014523735269904137, + 0.015160267241299152, + 0.00401215860620141, + 0.005966258700937033, + 0.004079489037394524, + 0.0046881381422281265, + 0.006299793254584074, + 0.005614349152892828, + 0.0063459621742367744, + 0.005441030953079462, + 0.017191994935274124, + 0.02861320972442627, + 0.021126171573996544, + 0.033266738057136536, + 0.030047403648495674, + 0.030408469960093498, + 0.014859683811664581, + 0.016671327874064445, + 0.00855288840830326, + 0.015319372527301311, + 0.01673080027103424, + 0.015405342914164066, + 0.012651175260543823, + 0.011003430932760239, + 0.0087931202724576, + 0.010279587469995022, + 0.005854408722370863, + 0.00700228288769722, + 0.007872541435062885, + 0.011787003837525845, + 0.0057490249164402485, + 0.005033064167946577, + 0.011489859782159328, + 0.00971389003098011, + 0.006951996125280857, + 0.010693765245378017, + 0.010174227878451347, + 0.010338779538869858, + 0.009067158214747906, + 0.0089447982609272, + 0.011104816570878029, + 0.007884526625275612, + 0.005724516697227955, + 0.007539215497672558, + 0.008870712481439114, + 0.00929170660674572, + 0.007961412891745567, + 0.005506649613380432, + 0.014634239487349987, + 0.0067528109066188335, + 0.00471640145406127, + 0.006626012269407511, + 0.007729851640760899, + 0.007671328727155924, + 0.010163639672100544, + 0.006576625630259514, + 0.004924682900309563, + 0.004030741285532713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.8983492851257324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134856224060059 + ], + "opposite_logprob_tokens": [ + -0.31348559260368347 + ], + "logprob_reward": -1.3134856224060059, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134856224060059, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24322885274887085, + 0.0329989455640316, + 0.029895588755607605, + 0.03029218502342701, + 0.0576779842376709, + 0.06133323162794113, + 0.039785727858543396, + 0.20094990730285645, + 0.051149703562259674, + 0.25268784165382385 + ], + "credit_map": [ + 0.01776117831468582, + 0.043993305414915085, + 0.034389834851026535, + 0.04213648661971092, + 0.04105338826775551, + 0.024997040629386902, + 0.016232609748840332, + 0.022665010765194893, + 0.0030461102724075317, + 0.004646316636353731, + 0.004194485489279032, + 0.004793843254446983, + 0.004578765016049147, + 0.00436973525211215, + 0.004170896019786596, + 0.0031987917609512806, + 0.0028222459368407726, + 0.0029009736608713865, + 0.003144562244415283, + 0.006064773537218571, + 0.0037484460044652224, + 0.0038955528289079666, + 0.0037453363183885813, + 0.0035736956633627415, + 0.002941504353657365, + 0.004892320837825537, + 0.004689710680395365, + 0.004317756742238998, + 0.0035656150430440903, + 0.0045877969823777676, + 0.0032921668607741594, + 0.0020053114276379347, + 0.005136460531502962, + 0.007880943827331066, + 0.009935169480741024, + 0.011139650829136372, + 0.007418584078550339, + 0.007300938479602337, + 0.004478603135794401, + 0.004387629218399525, + 0.005824035499244928, + 0.010618285275995731, + 0.011774969287216663, + 0.00879319291561842, + 0.00980575941503048, + 0.005156400613486767, + 0.004252810962498188, + 0.0051077790558338165, + 0.004129745531827211, + 0.0054306830279529095, + 0.007641462609171867, + 0.006735565606504679, + 0.004296501632779837, + 0.004134070128202438, + 0.0037775335367769003, + 0.003640167647972703, + 0.02094672992825508, + 0.02348717488348484, + 0.030553581193089485, + 0.03023453615605831, + 0.024230800569057465, + 0.023992890492081642, + 0.02536943554878235, + 0.02213476039469242, + 0.00493652606382966, + 0.007588586304336786, + 0.008502861484885216, + 0.006558452267199755, + 0.007800789084285498, + 0.00727880047634244, + 0.004585602320730686, + 0.0038980823010206223, + 0.020613452419638634, + 0.032844144850969315, + 0.025331754237413406, + 0.03920351713895798, + 0.04123485088348389, + 0.03060811199247837, + 0.041842859238386154, + 0.021009160205721855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -1.3134856224060059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "static wide-angle shot, shot from a level angle, captures the vast expanse of the ocean amidst a torrential downpour. A small wooden fishing boat, painted a faded red, stands as the sole focal point. Beneath a sky shrouded in thick, leaden clouds, a dense curtain of rain cascades down from the sky, completely blurring the line between sea and sky. The dark gray waters churn, whipping up surging waves capped with white foam. The boat appears vulnerable to this violent force. A massive wave lifts the bow almost perpendicular to the surface, only to slam it back into the trough, sending up a shower of white spray. The boat desperately sways from side to side, struggling to maintain its balance. The entire scene, bathed in cool, even, diffuse light, appears remarkably clean and crisp. The high shutter speed freezes every moment, making every splash and tumbling wave visible, devoid of motion blur. This extremely clear visual style contrasts sharply with the violent turbulence of the scene, highlighting the ruthless power of nature and the loneliness and tenacity of life struggling within it, creating a grand, cold and visually impactful atmosphere.", + "seed": 12717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_3636_seed_12717_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5460133906453848, + "vlm_logprob": -0.5460133906453848, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.5460133906453848, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10262300819158554 + ], + "opposite_logprob_tokens": [ + -2.9776229858398438 + ], + "logprob_reward": -0.10262300819158554, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.10262300819158554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12874439358711243, + 0.05203870311379433, + 0.1905461698770523, + 0.1901351660490036, + 0.05512794479727745, + 0.23647812008857727, + 0.055055081844329834, + 0.030436551198363304, + 0.027271071448922157, + 0.0341668538749218 + ], + "credit_map": [ + 0.006050784140825272, + 0.013957098126411438, + 0.010149607434868813, + 0.00699243089184165, + 0.03554756939411163, + 0.03498850017786026, + 0.011785145848989487, + 0.009273250587284565, + 0.0033476653043180704, + 0.006249609868973494, + 0.006370354909449816, + 0.009694048203527927, + 0.004893515259027481, + 0.008385604247450829, + 0.006150107830762863, + 0.006947795394808054, + 0.007651107851415873, + 0.04342466592788696, + 0.00793141033500433, + 0.008764935657382011, + 0.07687048614025116, + 0.03285800293087959, + 0.007367379032075405, + 0.005678168963640928, + 0.007669973652809858, + 0.06965666264295578, + 0.006289249286055565, + 0.01558717992156744, + 0.04785020276904106, + 0.020377829670906067, + 0.014399517327547073, + 0.008304520510137081, + 0.0037779968697577715, + 0.007975926622748375, + 0.008368096314370632, + 0.004282140638679266, + 0.007314991671591997, + 0.009465519338846207, + 0.006750696338713169, + 0.007192573510110378, + 0.014603495597839355, + 0.028059188276529312, + 0.03668221831321716, + 0.016651518642902374, + 0.03654586151242256, + 0.04175157472491264, + 0.020077545195817947, + 0.04210670292377472, + 0.004309383686631918, + 0.006225498393177986, + 0.007976938039064407, + 0.004393308889120817, + 0.009080038405954838, + 0.012568486854434013, + 0.005752462428063154, + 0.0047489614225924015, + 0.002196419285610318, + 0.002923286519944668, + 0.005574393086135387, + 0.006214061751961708, + 0.0035640376154333353, + 0.002789922757074237, + 0.0042048427276313305, + 0.0029695851262658834, + 0.0019981421064585447, + 0.005348339676856995, + 0.00355539214797318, + 0.002451216569170356, + 0.004412512760609388, + 0.004352598916739225, + 0.003270340384915471, + 0.0018825276056304574, + 0.002520647831261158, + 0.003896777518093586, + 0.0032772072590887547, + 0.0029085332062095404, + 0.00431183772161603, + 0.005614001769572496, + 0.007304101251065731, + 0.0043337419629096985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.10262300819158554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038108617067337036, + 0.0649845153093338, + 0.13157552480697632, + 0.11203287541866302, + 0.08666978776454926, + 0.1284930557012558, + 0.013323516584932804, + 0.011171674355864525, + 0.057693663984537125, + 0.35594677925109863 + ], + "credit_map": [ + 0.002866574563086033, + 0.00473306979984045, + 0.003853235859423876, + 0.008358646184206009, + 0.004862101282924414, + 0.004504133481532335, + 0.005335839465260506, + 0.003595016896724701, + 0.0026821347419172525, + 0.005475970450788736, + 0.0033405041322112083, + 0.006550964433699846, + 0.021291887387633324, + 0.017979880794882774, + 0.0047215670347213745, + 0.0029416107572615147, + 0.006410322152078152, + 0.019020814448595047, + 0.008367309346795082, + 0.0077262502163648605, + 0.04446442425251007, + 0.029610024765133858, + 0.007930872030556202, + 0.008045508526265621, + 0.004222309682518244, + 0.031140968203544617, + 0.005257876589894295, + 0.017956314608454704, + 0.02856653556227684, + 0.007505691610276699, + 0.010002510622143745, + 0.007380667608231306, + 0.0055487314239144325, + 0.009471322409808636, + 0.013087233528494835, + 0.013628363609313965, + 0.011001894250512123, + 0.011173009872436523, + 0.008624409325420856, + 0.01413482055068016, + 0.008192040957510471, + 0.015131459571421146, + 0.01294679380953312, + 0.0064070564694702625, + 0.02729109115898609, + 0.03043052740395069, + 0.011343685910105705, + 0.01675039529800415, + 0.000959340832196176, + 0.001400528009980917, + 0.0019230119651183486, + 0.0011794340098276734, + 0.0013831352116540074, + 0.0017707730876281857, + 0.002125656930729747, + 0.002581637119874358, + 0.000923969957511872, + 0.0015289860311895609, + 0.002104399725794792, + 0.0011990648927167058, + 0.001496333396062255, + 0.0014809692511335015, + 0.0011964533478021622, + 0.0012414976954460144, + 0.0034160022623836994, + 0.01180953811854124, + 0.005887933541089296, + 0.005795533303171396, + 0.006165512837469578, + 0.010976445861160755, + 0.009872539900243282, + 0.0037701574619859457, + 0.023724233731627464, + 0.027524331584572792, + 0.039230309426784515, + 0.03643427789211273, + 0.03925766423344612, + 0.05191406235098839, + 0.10147924721240997, + 0.036382630467414856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002592831850052 + ], + "opposite_logprob_tokens": [ + -2.69502592086792 + ], + "logprob_reward": -0.07002592831850052, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002592831850052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3704463243484497, + 0.052666544914245605, + 0.17032378911972046, + 0.0764685794711113, + 0.06479071825742722, + 0.09918751567602158, + 0.02785903960466385, + 0.03754040598869324, + 0.03406856209039688, + 0.06664853543043137 + ], + "credit_map": [ + 0.018711892887949944, + 0.03568565845489502, + 0.021517885848879814, + 0.014772686176002026, + 0.12262006849050522, + 0.10935428738594055, + 0.02654639258980751, + 0.021237432956695557, + 0.0033747719135135412, + 0.007018832489848137, + 0.00428490387275815, + 0.01180193480104208, + 0.006903768051415682, + 0.00659619364887476, + 0.008011755533516407, + 0.004674385767430067, + 0.005107930861413479, + 0.04477110877633095, + 0.005275574047118425, + 0.009785911068320274, + 0.07166537642478943, + 0.025282228365540504, + 0.004716495051980019, + 0.0037191668525338173, + 0.0038770302198827267, + 0.020208517089486122, + 0.009173877537250519, + 0.006336268503218889, + 0.013640216551721096, + 0.009103884920477867, + 0.004941259976476431, + 0.009187519550323486, + 0.0033798227086663246, + 0.011307905428111553, + 0.00828290730714798, + 0.004975086078047752, + 0.010675660334527493, + 0.01254713162779808, + 0.00680297939106822, + 0.006819228641688824, + 0.0076357778161764145, + 0.0128223467618227, + 0.014430388808250427, + 0.006964309141039848, + 0.009836671873927116, + 0.010554083622992039, + 0.009772127494215965, + 0.027171814814209938, + 0.002541020279750228, + 0.0032651524525135756, + 0.005634813103824854, + 0.002728796564042568, + 0.0033590078819543123, + 0.0031209716107696295, + 0.004398868884891272, + 0.002810406731441617, + 0.0024849006440490484, + 0.004037379752844572, + 0.002925147535279393, + 0.0047896187752485275, + 0.008142372593283653, + 0.0056945257820189, + 0.006174842827022076, + 0.0032916199415922165, + 0.0025760121643543243, + 0.006244876887649298, + 0.005165007431060076, + 0.003529411507770419, + 0.0036427939776331186, + 0.006995071191340685, + 0.00332837481983006, + 0.0025870157405734062, + 0.004239540081471205, + 0.00597790302708745, + 0.0065108416602015495, + 0.006553133949637413, + 0.008362666703760624, + 0.008851578459143639, + 0.0162547267973423, + 0.009898144751787186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.07002592831850052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328458189964294 + ], + "opposite_logprob_tokens": [ + -0.7578458189964294 + ], + "logprob_reward": -0.6328458189964294, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328458189964294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16438646614551544, + 0.08521672338247299, + 0.13698996603488922, + 0.13324877619743347, + 0.0727936178445816, + 0.19995421171188354, + 0.025398001074790955, + 0.060282230377197266, + 0.08021855354309082, + 0.04151146113872528 + ], + "credit_map": [ + 0.010661562904715538, + 0.030506467446684837, + 0.015505186282098293, + 0.016483495011925697, + 0.022169427946209908, + 0.039301324635744095, + 0.015945812687277794, + 0.013813182711601257, + 0.004079235251992941, + 0.011305015534162521, + 0.004259440582245588, + 0.007207693997770548, + 0.028298677876591682, + 0.018035631626844406, + 0.007001377642154694, + 0.0050296541303396225, + 0.006638134364038706, + 0.018495768308639526, + 0.00921960361301899, + 0.01054079458117485, + 0.04373260214924812, + 0.03081243857741356, + 0.008399365469813347, + 0.009151255711913109, + 0.0062567489221692085, + 0.03732035309076309, + 0.008979672566056252, + 0.013764708302915096, + 0.029760921373963356, + 0.011279609985649586, + 0.016070978716015816, + 0.009815778583288193, + 0.004835794679820538, + 0.008848320692777634, + 0.0116672832518816, + 0.011352558620274067, + 0.00906135980039835, + 0.006805181968957186, + 0.009172899648547173, + 0.011050218716263771, + 0.017396723851561546, + 0.0249954666942358, + 0.025644345209002495, + 0.015097888186573982, + 0.027793938294053078, + 0.02085166610777378, + 0.01791459135711193, + 0.05025959759950638, + 0.0018719962099567056, + 0.002913772128522396, + 0.004324256908148527, + 0.002592843258753419, + 0.0034991223365068436, + 0.0029548322781920433, + 0.004765222780406475, + 0.002475955756381154, + 0.00421055406332016, + 0.0065422384068369865, + 0.00618441216647625, + 0.007751777768135071, + 0.01218830794095993, + 0.010443436913192272, + 0.004744936246424913, + 0.008216568268835545, + 0.005588924512267113, + 0.015449381433427334, + 0.00845385529100895, + 0.007802021689713001, + 0.008928273804485798, + 0.01403532549738884, + 0.01238991878926754, + 0.007570852059870958, + 0.0033080156426876783, + 0.00414696242660284, + 0.004204859491437674, + 0.004334828816354275, + 0.006041449494659901, + 0.006207217462360859, + 0.0076489015482366085, + 0.005619225092232227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.6328458189964294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103596210479736 + ], + "opposite_logprob_tokens": [ + -0.16035957634449005 + ], + "logprob_reward": -1.9103596210479736, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103596210479736, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24505700170993805, + 0.08657470345497131, + 0.11675678938627243, + 0.10114949196577072, + 0.04779773950576782, + 0.1298215389251709, + 0.054373886436223984, + 0.09920202195644379, + 0.029831375926733017, + 0.0894353836774826 + ], + "credit_map": [ + 0.014853265136480331, + 0.050058815628290176, + 0.02056935243308544, + 0.029444463551044464, + 0.05392741411924362, + 0.029941251501441002, + 0.02485533244907856, + 0.02140713669359684, + 0.005980886053293943, + 0.016652068123221397, + 0.005710679106414318, + 0.012437701225280762, + 0.01567184180021286, + 0.016023952513933182, + 0.007510453928261995, + 0.006587127223610878, + 0.0041665988974273205, + 0.01508831512182951, + 0.008296428248286247, + 0.008109218440949917, + 0.03743552416563034, + 0.02669544331729412, + 0.008563973940908909, + 0.008401294238865376, + 0.008589598350226879, + 0.01144611556082964, + 0.012045104056596756, + 0.014669199474155903, + 0.01340960431843996, + 0.014725525863468647, + 0.013925908133387566, + 0.01233843993395567, + 0.0036210448015481234, + 0.005091026891022921, + 0.010371603071689606, + 0.006909054238349199, + 0.004706174600869417, + 0.005915469955652952, + 0.005333912093192339, + 0.005849460605531931, + 0.011476296000182629, + 0.016653060913085938, + 0.017084963619709015, + 0.0090712271630764, + 0.019000327214598656, + 0.017778748646378517, + 0.01147688739001751, + 0.027280034497380257, + 0.005310813430696726, + 0.006666047498583794, + 0.010015984065830708, + 0.0041418615728616714, + 0.00903542060405016, + 0.007587081287056208, + 0.005454628262668848, + 0.006162054371088743, + 0.006677535362541676, + 0.011236085556447506, + 0.00859006680548191, + 0.010480858385562897, + 0.02120075188577175, + 0.017503028735518456, + 0.013092107139527798, + 0.010421596467494965, + 0.0025938458275049925, + 0.003426269395276904, + 0.004720022901892662, + 0.0038850009441375732, + 0.00369632919318974, + 0.0037243186961859465, + 0.0044186534360051155, + 0.0033669378608465195, + 0.006633951794356108, + 0.008624261245131493, + 0.013650047592818737, + 0.01085616648197174, + 0.011910872533917427, + 0.011185908690094948, + 0.01766107976436615, + 0.008913093246519566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -1.9103596210479736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static wide-angle shot, shot from a level angle, captures the vast expanse of the ocean amidst a torrential downpour. A small wooden fishing boat, painted a faded red, stands as the sole focal point. Beneath a sky shrouded in thick, leaden clouds, a dense curtain of rain cascades down from the sky, completely blurring the line between sea and sky. The dark gray waters churn, whipping up surging waves capped with white foam. The boat appears vulnerable to this violent force. A massive wave lifts the bow almost perpendicular to the surface, only to slam it back into the trough, sending up a shower of white spray. The boat desperately sways from side to side, struggling to maintain its balance. The entire scene, bathed in cool, even, diffuse light, appears remarkably clean and crisp. The high shutter speed freezes every moment, making every splash and tumbling wave visible, devoid of motion blur. This extremely clear visual style contrasts sharply with the violent turbulence of the scene, highlighting the ruthless power of nature and the loneliness and tenacity of life struggling within it, creating a grand, cold and visually impactful atmosphere.", + "seed": 12718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_3636_seed_12718_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.47684098072350023, + "vlm_logprob": -0.47684098072350023, + "vlm_yes_no_margin": 1.5, + "vlm_token_credit": -0.47684098072350023, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14153490960597992 + ], + "opposite_logprob_tokens": [ + -2.3915348052978516 + ], + "logprob_reward": -0.14153490960597992, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.14153490960597992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06129206344485283, + 0.033454764634370804, + 0.1899804323911667, + 0.027127280831336975, + 0.0380583256483078, + 0.07602148503065109, + 0.058933112770318985, + 0.2597489058971405, + 0.09633706510066986, + 0.15904657542705536 + ], + "credit_map": [ + 0.003761052852496505, + 0.009262190200388432, + 0.0077607398852705956, + 0.006304549518972635, + 0.011767840012907982, + 0.008916804566979408, + 0.007858699187636375, + 0.005660186987370253, + 0.002087110187858343, + 0.007149597629904747, + 0.004084107466042042, + 0.0020077775698155165, + 0.010589322075247765, + 0.0042780558578670025, + 0.001312950742430985, + 0.001945843338035047, + 0.005889770574867725, + 0.060915879905223846, + 0.005208164919167757, + 0.003378338413313031, + 0.02436494641005993, + 0.08216899633407593, + 0.0024111007805913687, + 0.005643236916512251, + 0.001903385273180902, + 0.0040846155025064945, + 0.0042562345042824745, + 0.0032616090029478073, + 0.0042482889257371426, + 0.00313738570548594, + 0.0024168756790459156, + 0.0038188868202269077, + 0.002859072759747505, + 0.006646167952567339, + 0.005552488379180431, + 0.003946466837078333, + 0.003993733320385218, + 0.007869953289628029, + 0.0028525537345558405, + 0.004337887745350599, + 0.00440563540905714, + 0.020821385085582733, + 0.009569087065756321, + 0.00729356100782752, + 0.009682954289019108, + 0.012767639011144638, + 0.006950523238629103, + 0.0045307050459086895, + 0.003070783568546176, + 0.006472985725849867, + 0.005119935143738985, + 0.003951528109610081, + 0.0134825324639678, + 0.018085263669490814, + 0.0044678375124931335, + 0.004282251000404358, + 0.0065883793868124485, + 0.11104399710893631, + 0.0017534083453938365, + 0.0028880301397293806, + 0.11802320182323456, + 0.014283835887908936, + 0.0028516652528196573, + 0.0023164106532931328, + 0.00714479573071003, + 0.02030504308640957, + 0.013382062315940857, + 0.010847184807062149, + 0.012449684552848339, + 0.013811024837195873, + 0.008591901510953903, + 0.009805372916162014, + 0.005979166366159916, + 0.05303310975432396, + 0.008532441221177578, + 0.010181738063693047, + 0.057625554502010345, + 0.011722717434167862, + 0.006152112502604723, + 0.00581973185762763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.14153490960597992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822619631886482 + ], + "opposite_logprob_tokens": [ + -3.529822587966919 + ], + "logprob_reward": -0.029822619631886482, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822619631886482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18126431107521057, + 0.05482078343629837, + 0.21889396011829376, + 0.01629018597304821, + 0.014701992273330688, + 0.014561428688466549, + 0.029965287074446678, + 0.4351329207420349, + 0.017318597063422203, + 0.017050527036190033 + ], + "credit_map": [ + 0.011337067931890488, + 0.02517264150083065, + 0.027322575449943542, + 0.0237825196236372, + 0.028830328956246376, + 0.016810933127999306, + 0.029113784432411194, + 0.018894465640187263, + 0.00308767007663846, + 0.011404736898839474, + 0.004295704420655966, + 0.0027709712740033865, + 0.017638033255934715, + 0.006638386286795139, + 0.004833894316107035, + 0.004151382949203253, + 0.006397812627255917, + 0.09912187606096268, + 0.00391268078237772, + 0.003152439370751381, + 0.02618960104882717, + 0.07157290726900101, + 0.004012365825474262, + 0.0045342762023210526, + 0.0012998977908864617, + 0.002014144789427519, + 0.0017317216843366623, + 0.0019312080694362521, + 0.002811306156218052, + 0.0026659162249416113, + 0.0018623712239786983, + 0.0019736196845769882, + 0.0009022956364788115, + 0.002625242806971073, + 0.0018715813057497144, + 0.0017295180587098002, + 0.0014216422569006681, + 0.002547775860875845, + 0.0018981280736625195, + 0.0017058091470971704, + 0.000957221316639334, + 0.002034858101978898, + 0.0025184908881783485, + 0.0017517979722470045, + 0.0024302892852574587, + 0.0019053831929340959, + 0.0015384267317131162, + 0.0014249617233872414, + 0.0012384954607114196, + 0.005855845753103495, + 0.002488853642717004, + 0.0028856820426881313, + 0.004199418239295483, + 0.008629659190773964, + 0.00221750745549798, + 0.0024498228449374437, + 0.011095099151134491, + 0.182530477643013, + 0.0031664008274674416, + 0.0049433521926403046, + 0.1979600489139557, + 0.02481280267238617, + 0.005915745627135038, + 0.004708991385996342, + 0.00121108943130821, + 0.0036606602370738983, + 0.0023171219509094954, + 0.0018868447514250875, + 0.00240575079806149, + 0.0026595161762088537, + 0.0016760096186771989, + 0.0015016039833426476, + 0.0012427917681634426, + 0.0031925721559673548, + 0.0020429077558219433, + 0.0016132270684465766, + 0.0036349960137158632, + 0.0018513716058805585, + 0.0012999068712815642, + 0.002172753680497408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.029822619631886482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656366765499115 + ], + "opposite_logprob_tokens": [ + -3.048656463623047 + ], + "logprob_reward": -0.048656366765499115, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656366765499115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02375919371843338, + 0.03640887141227722, + 0.1285446286201477, + 0.025829622521996498, + 0.015300732105970383, + 0.02074260264635086, + 0.01206586230546236, + 0.6763350963592529, + 0.01596860960125923, + 0.04504477232694626 + ], + "credit_map": [ + 0.001565135084092617, + 0.005146340932697058, + 0.002343946136534214, + 0.0014855278423056006, + 0.0051598153077065945, + 0.004026321694254875, + 0.001508216722868383, + 0.0025238902308046818, + 0.001711096498183906, + 0.005988306365907192, + 0.0015537114813923836, + 0.0009751280304044485, + 0.014740665443241596, + 0.00824876967817545, + 0.00125744822435081, + 0.001933744177222252, + 0.0038001195061951876, + 0.058132853358983994, + 0.0029737967997789383, + 0.0018933165119960904, + 0.01042788103222847, + 0.048213087022304535, + 0.0012559150345623493, + 0.0018476671539247036, + 0.0019399209413677454, + 0.005352579057216644, + 0.002946116030216217, + 0.0022326072212308645, + 0.0054869684390723705, + 0.003601162461563945, + 0.0013177369255572557, + 0.0029525321442633867, + 0.0010323509341105819, + 0.0019203861011192203, + 0.0019355028634890914, + 0.0024369440507143736, + 0.002561999252066016, + 0.0020093268249183893, + 0.0018814884824678302, + 0.0015227339463308454, + 0.0009563020430505276, + 0.007404400035738945, + 0.001138303312472999, + 0.0011161845177412033, + 0.006102239713072777, + 0.0017878207145258784, + 0.0015892103547230363, + 0.0006481417804025114, + 0.0006540737813338637, + 0.0013484583469107747, + 0.0014973394572734833, + 0.0010958572383970022, + 0.002323956461623311, + 0.001581162796355784, + 0.0017074727220460773, + 0.0018575417343527079, + 0.017268020659685135, + 0.28872889280319214, + 0.006360153201967478, + 0.008181074634194374, + 0.3051328957080841, + 0.03747676685452461, + 0.007065265905112028, + 0.0061220405623316765, + 0.000665563449729234, + 0.005392917897552252, + 0.000704804842825979, + 0.0007467783871106803, + 0.005177557002753019, + 0.0011564885498955846, + 0.0014342897338792682, + 0.0006902092136442661, + 0.0017590454081073403, + 0.014125710353255272, + 0.00243009184487164, + 0.002815164625644684, + 0.015207265503704548, + 0.003282260149717331, + 0.0020984329748898745, + 0.003326802747324109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.048656366765499115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578279972076416 + ], + "opposite_logprob_tokens": [ + -0.6328279972076416 + ], + "logprob_reward": -0.7578279972076416, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578279972076416, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02568291500210762, + 0.01908029243350029, + 0.25726309418678284, + 0.010713216848671436, + 0.007249315269291401, + 0.02733345702290535, + 0.010436720214784145, + 0.5899307727813721, + 0.03519253060221672, + 0.01711767353117466 + ], + "credit_map": [ + 0.0017565853195264935, + 0.00378058897331357, + 0.0036378137301653624, + 0.003377515822649002, + 0.003685321658849716, + 0.002678957534953952, + 0.0039299349300563335, + 0.002836195519194007, + 0.0009127211524173617, + 0.004232720006257296, + 0.00127867900300771, + 0.0006147882086224854, + 0.007478700485080481, + 0.0030293844174593687, + 0.0007185912691056728, + 0.0008147077751345932, + 0.0069583808071911335, + 0.11932606995105743, + 0.002828559372574091, + 0.00263189896941185, + 0.018316790461540222, + 0.10221995413303375, + 0.002001883229240775, + 0.0029795595910400152, + 0.0006442730664275587, + 0.001709837932139635, + 0.0009436922264285386, + 0.0008091395138762891, + 0.0017853748286142945, + 0.0031720465049147606, + 0.0006681233644485474, + 0.0009807287715375423, + 0.0005433256737887859, + 0.0012611030833795667, + 0.0008034489583224058, + 0.0009909872896969318, + 0.0010969602735713124, + 0.0009537426522001624, + 0.0008392198942601681, + 0.0007605274440720677, + 0.0014514954527840018, + 0.00895414873957634, + 0.002342689549550414, + 0.0019886880181729794, + 0.004216825123876333, + 0.004701073281466961, + 0.002249479992315173, + 0.001429057097993791, + 0.0008534312946721911, + 0.0015301575185731053, + 0.0012896191328763962, + 0.0011205264599993825, + 0.001739120576530695, + 0.0012355876388028264, + 0.0012282239040359855, + 0.0014400536892935634, + 0.014952342957258224, + 0.24999099969863892, + 0.004110788926482201, + 0.006549189332872629, + 0.26840829849243164, + 0.033311355859041214, + 0.00703104492276907, + 0.005576772149652243, + 0.0020737950690090656, + 0.008794252760708332, + 0.004213424399495125, + 0.0036166184581816196, + 0.006980287376791239, + 0.003634322201833129, + 0.002851600293070078, + 0.00302823050878942, + 0.0007680393173359334, + 0.005806322675198317, + 0.0012399032711982727, + 0.0010704232845455408, + 0.005492870230227709, + 0.0012519892770797014, + 0.0007388780941255391, + 0.000749247963540256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.7578279972076416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063630104064941 + ], + "opposite_logprob_tokens": [ + -0.28136301040649414 + ], + "logprob_reward": -1.4063630104064941, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063630104064941, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03720724582672119, + 0.03711071237921715, + 0.2675454914569855, + 0.011319153010845184, + 0.018012329936027527, + 0.0204547680914402, + 0.032821737229824066, + 0.29178711771965027, + 0.17348405718803406, + 0.1102573499083519 + ], + "credit_map": [ + 0.0027057554107159376, + 0.00474447151646018, + 0.008251889608800411, + 0.004342496860772371, + 0.004531217273324728, + 0.005186948925256729, + 0.003481325227767229, + 0.003963141236454248, + 0.0015539657324552536, + 0.0071787722408771515, + 0.0017393393209204078, + 0.0011400280054658651, + 0.015354231931269169, + 0.0077646574936807156, + 0.0010693889344111085, + 0.0013103282544761896, + 0.007240258622914553, + 0.12294459342956543, + 0.0036088970955461264, + 0.003715748665854335, + 0.020670661702752113, + 0.10327430069446564, + 0.002311059972271323, + 0.0037799892015755177, + 0.0010663603898137808, + 0.0014881069073453546, + 0.0017779801273718476, + 0.0012137416051700711, + 0.0018223437946289778, + 0.0015517591964453459, + 0.0010935446480289102, + 0.00130531657487154, + 0.0013774551916867495, + 0.0029061981476843357, + 0.001854557660408318, + 0.001584756770171225, + 0.0039156461134552956, + 0.003422385547310114, + 0.0013409815728664398, + 0.0016103490488603711, + 0.001310658873990178, + 0.0059441509656608105, + 0.0020094774663448334, + 0.0017079419922083616, + 0.004156035371124744, + 0.0022772536613047123, + 0.0018232223810628057, + 0.0012260270304977894, + 0.002634167205542326, + 0.005562158301472664, + 0.003927574958652258, + 0.0026788567192852497, + 0.008342096582055092, + 0.0036874280776828527, + 0.0034395705442875624, + 0.0025498869363218546, + 0.007293910253793001, + 0.12419643998146057, + 0.002235274063423276, + 0.0030250530689954758, + 0.13271625339984894, + 0.016410723328590393, + 0.00327893509529531, + 0.002630502451211214, + 0.012112385593354702, + 0.03780960291624069, + 0.024253394454717636, + 0.019855795428156853, + 0.026956724002957344, + 0.02156110294163227, + 0.014800412580370903, + 0.01613464578986168, + 0.004364331718534231, + 0.03747240826487541, + 0.005984905641525984, + 0.005877081770449877, + 0.0395662747323513, + 0.00903477892279625, + 0.004088874906301498, + 0.0038686885964125395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -1.4063630104064941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "static wide-angle shot, shot from a level angle, captures the vast expanse of the ocean amidst a torrential downpour. A small wooden fishing boat, painted a faded red, stands as the sole focal point. Beneath a sky shrouded in thick, leaden clouds, a dense curtain of rain cascades down from the sky, completely blurring the line between sea and sky. The dark gray waters churn, whipping up surging waves capped with white foam. The boat appears vulnerable to this violent force. A massive wave lifts the bow almost perpendicular to the surface, only to slam it back into the trough, sending up a shower of white spray. The boat desperately sways from side to side, struggling to maintain its balance. The entire scene, bathed in cool, even, diffuse light, appears remarkably clean and crisp. The high shutter speed freezes every moment, making every splash and tumbling wave visible, devoid of motion blur. This extremely clear visual style contrasts sharply with the violent turbulence of the scene, highlighting the ruthless power of nature and the loneliness and tenacity of life struggling within it, creating a grand, cold and visually impactful atmosphere.", + "seed": 12719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_14/index_3636_seed_12719_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5447812769562006, + "vlm_logprob": -0.5447812769562006, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.5447812769562006, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14920079708099365 + ], + "opposite_logprob_tokens": [ + -2.274200916290283 + ], + "logprob_reward": -0.14920079708099365, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.14920079708099365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047878798097372055, + 0.026135236024856567, + 0.012634072452783585, + 0.11110736429691315, + 0.23173995316028595, + 0.3557717204093933, + 0.137709379196167, + 0.02065822295844555, + 0.026431048288941383, + 0.029934249818325043 + ], + "credit_map": [ + 0.0022228537127375603, + 0.010852513834834099, + 0.004916639998555183, + 0.004906446672976017, + 0.003818490542471409, + 0.01274607703089714, + 0.0029516902286559343, + 0.0054640863090753555, + 0.0014650976518169045, + 0.003786026732996106, + 0.0028457846492528915, + 0.002934470772743225, + 0.005629586521536112, + 0.004115658346563578, + 0.0026323285419493914, + 0.002726282924413681, + 0.0008728725952096283, + 0.002196753164753318, + 0.0011132616782560945, + 0.0012211232678964734, + 0.002554690930992365, + 0.0015216837637126446, + 0.0009729480370879173, + 0.00218073814176023, + 0.003250514855608344, + 0.03959265351295471, + 0.0035484526306390762, + 0.0020128667820245028, + 0.048051103949546814, + 0.007971351966261864, + 0.001730988617055118, + 0.0049494365230202675, + 0.0061008259654045105, + 0.0881223976612091, + 0.006616772152483463, + 0.006436673458665609, + 0.0982166975736618, + 0.011546493507921696, + 0.004946068394929171, + 0.009754016064107418, + 0.009002587758004665, + 0.1389244943857193, + 0.005979918409138918, + 0.0056267548352479935, + 0.16187766194343567, + 0.02078366093337536, + 0.007097756955772638, + 0.006478870753198862, + 0.007450864650309086, + 0.04079391062259674, + 0.01647000201046467, + 0.007492770440876484, + 0.034599609673023224, + 0.010205533355474472, + 0.009269514121115208, + 0.011427167803049088, + 0.0019078842597082257, + 0.0032663068268448114, + 0.003326910547912121, + 0.001660723239183426, + 0.0038209527265280485, + 0.002545646857470274, + 0.0015754976775497198, + 0.0025543011724948883, + 0.0014720129547640681, + 0.0038630524650216103, + 0.0027226500678807497, + 0.002836523111909628, + 0.005600395612418652, + 0.003952840808779001, + 0.0029504455160349607, + 0.003033128334209323, + 0.002339953789487481, + 0.0054984926246106625, + 0.0033617662265896797, + 0.002817227505147457, + 0.005956875625997782, + 0.0030588123481720686, + 0.004063431639224291, + 0.0028376909904181957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.14920079708099365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026372551918029785 + ], + "opposite_logprob_tokens": [ + -3.6513724327087402 + ], + "logprob_reward": -0.026372551918029785, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026372551918029785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03913847729563713, + 0.15618565678596497, + 0.02292712964117527, + 0.14670196175575256, + 0.11010795086622238, + 0.22225260734558105, + 0.18421298265457153, + 0.03294843062758446, + 0.057693369686603546, + 0.027831347659230232 + ], + "credit_map": [ + 0.0029125006403774023, + 0.005040066782385111, + 0.00610025180503726, + 0.004512031562626362, + 0.006896170321851969, + 0.00509232934564352, + 0.004086196422576904, + 0.004498932510614395, + 0.00970916636288166, + 0.018106132745742798, + 0.019035879522562027, + 0.01105785183608532, + 0.035368531942367554, + 0.03804133087396622, + 0.0113626504316926, + 0.013504133559763432, + 0.0016402560286223888, + 0.00309359235689044, + 0.002044335938990116, + 0.004254580475389957, + 0.004043239634484053, + 0.00259559229016304, + 0.0021086186170578003, + 0.003146916162222624, + 0.0044220490381121635, + 0.05934632569551468, + 0.003172420896589756, + 0.006070992909371853, + 0.060360461473464966, + 0.007186349481344223, + 0.0033154760021716356, + 0.0028279083780944347, + 0.0025746391620486975, + 0.041760411113500595, + 0.0031658492516726255, + 0.0025860026944428682, + 0.0479155108332634, + 0.006357198115438223, + 0.004027397837489843, + 0.0017209487268701196, + 0.005405402276664972, + 0.09014178812503815, + 0.003937344066798687, + 0.004208861850202084, + 0.09987697750329971, + 0.010313788428902626, + 0.004844986833631992, + 0.0035234689712524414, + 0.005139960441738367, + 0.07215455174446106, + 0.009199334308505058, + 0.004159615375101566, + 0.07312662899494171, + 0.007948197424411774, + 0.005964041221886873, + 0.006520672235637903, + 0.0019214838976040483, + 0.003564021782949567, + 0.002578088780865073, + 0.003691338235512376, + 0.008078528568148613, + 0.006524307187646627, + 0.003865413833409548, + 0.0027252486906945705, + 0.0038546763826161623, + 0.01144428737461567, + 0.005387571640312672, + 0.009832318872213364, + 0.010355949401855469, + 0.004675491247326136, + 0.006046219263225794, + 0.006096858065575361, + 0.0019220199901610613, + 0.004442613106220961, + 0.002422486199066043, + 0.0023686918430030346, + 0.0074281590059399605, + 0.0033668966498225927, + 0.003962344024330378, + 0.0019181407988071442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.026372551918029785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588146522641182 + ], + "opposite_logprob_tokens": [ + -3.8955881595611572 + ], + "logprob_reward": -0.020588146522641182, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588146522641182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04086471349000931, + 0.024593932554125786, + 0.04517350345849991, + 0.4394516944885254, + 0.17707712948322296, + 0.030828576534986496, + 0.09600276499986649, + 0.01717168278992176, + 0.05510254204273224, + 0.07373346388339996 + ], + "credit_map": [ + 0.0025036416482180357, + 0.009437933564186096, + 0.004188437946140766, + 0.00300432275980711, + 0.0038685158360749483, + 0.009890791960060596, + 0.002290853066369891, + 0.005680217407643795, + 0.001802746206521988, + 0.002756045199930668, + 0.004755838308483362, + 0.0030580705497413874, + 0.0028778475243598223, + 0.003729656571522355, + 0.00216915481723845, + 0.003444576170295477, + 0.0023761973716318607, + 0.010672276839613914, + 0.0014013011241331697, + 0.0023808921687304974, + 0.015191532671451569, + 0.0061081089079380035, + 0.002417762763798237, + 0.004625432658940554, + 0.010394244454801083, + 0.17576417326927185, + 0.007575775496661663, + 0.0047744400799274445, + 0.19844026863574982, + 0.025178952142596245, + 0.0043922606855630875, + 0.01293154712766409, + 0.004038160666823387, + 0.0658552423119545, + 0.005822868086397648, + 0.0042852433398365974, + 0.0766730010509491, + 0.010986939072608948, + 0.00658427132293582, + 0.002831409452483058, + 0.0011886615538969636, + 0.008681141771376133, + 0.001736898673698306, + 0.0013526978436857462, + 0.01223165262490511, + 0.0029945108108222485, + 0.0013763020979240537, + 0.001266710227355361, + 0.002680879319086671, + 0.0371377132833004, + 0.0024523905012756586, + 0.0014043349074199796, + 0.04331371560692787, + 0.005735875107347965, + 0.0015943621983751655, + 0.0016834995476529002, + 0.0008746612002141774, + 0.004073591437190771, + 0.0009190047858282924, + 0.0020610939245671034, + 0.004622656852006912, + 0.0013701050775125623, + 0.0023676021955907345, + 0.0008829673170112073, + 0.003932449501007795, + 0.011173911392688751, + 0.005838122684508562, + 0.006311747245490551, + 0.011076771654188633, + 0.0059685963205993176, + 0.004958733916282654, + 0.0058422149159014225, + 0.005140305031090975, + 0.010262400843203068, + 0.006599137093871832, + 0.007705394644290209, + 0.0153583325445652, + 0.009004502557218075, + 0.013641098514199257, + 0.006022295448929071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.020588146522641182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261755704879761 + ], + "opposite_logprob_tokens": [ + -0.5761755704879761 + ], + "logprob_reward": -0.8261755704879761, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261755704879761, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04203013330698013, + 0.06335870921611786, + 0.03734618052840233, + 0.5379291772842407, + 0.035663578659296036, + 0.09540604054927826, + 0.10759742558002472, + 0.03186239302158356, + 0.022268587723374367, + 0.026537766680121422 + ], + "credit_map": [ + 0.002772984327748418, + 0.0053385295905172825, + 0.006959361489862204, + 0.004754561465233564, + 0.006226162426173687, + 0.007995431311428547, + 0.0034747186582535505, + 0.004508384969085455, + 0.0039460305124521255, + 0.007247606758028269, + 0.007922768592834473, + 0.00613007228821516, + 0.014104298315942287, + 0.014911040663719177, + 0.0038660946302115917, + 0.005230795592069626, + 0.002273169346153736, + 0.008478031493723392, + 0.0027006512973457575, + 0.002826876472681761, + 0.00848691537976265, + 0.004394940100610256, + 0.002176393289119005, + 0.00600920245051384, + 0.012429933063685894, + 0.22140710055828094, + 0.005336314905434847, + 0.004292786121368408, + 0.2534876763820648, + 0.031140994280576706, + 0.00456662755459547, + 0.005267776548862457, + 0.002134727081283927, + 0.005334038287401199, + 0.005203505512326956, + 0.00507163954898715, + 0.0043466598726809025, + 0.003126284806057811, + 0.006239483132958412, + 0.004207238554954529, + 0.0024075638502836227, + 0.039098821580410004, + 0.0012323311530053616, + 0.001025979407131672, + 0.044705212116241455, + 0.005110860336571932, + 0.0011000025551766157, + 0.0007252729847095907, + 0.002707752166315913, + 0.04379863664507866, + 0.0024583004415035248, + 0.0014556209789589047, + 0.04864345118403435, + 0.00534913782030344, + 0.0015124422498047352, + 0.0016720861895009875, + 0.0021760554518550634, + 0.005599938798695803, + 0.0043794261291623116, + 0.0029825777746737003, + 0.006063599605113268, + 0.004178720526397228, + 0.003275024238973856, + 0.003207048401236534, + 0.0015252341981977224, + 0.004667247179895639, + 0.002984334249049425, + 0.0027114003896713257, + 0.003731467528268695, + 0.0019160025985911489, + 0.0018824029248207808, + 0.002850499004125595, + 0.0017805756069719791, + 0.0038247474003583193, + 0.002355023054406047, + 0.0028846519999206066, + 0.0048077902756631374, + 0.0035378988832235336, + 0.005137060768902302, + 0.0022100156638771296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.8261755704879761 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static wide-angle view taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden fishing boat painted a faded red visible as the sole focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a massive wave lifting the boat's bow almost perpendicular to the surface before slamming it back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boat visibly swaying from side to side while struggling to maintain balance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style extremely clear and crisp with no motion blur, freezing every splash and wave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015693187713623 + ], + "opposite_logprob_tokens": [ + -0.20156925916671753 + ], + "logprob_reward": -1.7015693187713623, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015693187713623, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10078532248735428, + 0.08148019760847092, + 0.06227106600999832, + 0.3995361328125, + 0.09259703755378723, + 0.15860360860824585, + 0.018809638917446136, + 0.030077550560235977, + 0.026628222316503525, + 0.029211126267910004 + ], + "credit_map": [ + 0.006222087424248457, + 0.020047837868332863, + 0.010288967750966549, + 0.01215039100497961, + 0.021395087242126465, + 0.01186366192996502, + 0.00881874468177557, + 0.00999856274574995, + 0.004372060764580965, + 0.011296556331217289, + 0.007075057830661535, + 0.011765506118535995, + 0.017389539629220963, + 0.013599615544080734, + 0.007667484227567911, + 0.008314387872815132, + 0.002812315709888935, + 0.015588073991239071, + 0.0028972772415727377, + 0.003160124411806464, + 0.020966362208127975, + 0.0060651302337646484, + 0.002568730153143406, + 0.008213055320084095, + 0.00979882013052702, + 0.15687203407287598, + 0.0059862588532269, + 0.005041921511292458, + 0.18377749621868134, + 0.02671387605369091, + 0.00446444982662797, + 0.006881335750222206, + 0.004420898389071226, + 0.025054600089788437, + 0.005178977269679308, + 0.007132441736757755, + 0.031596578657627106, + 0.0080023892223835, + 0.0071087186224758625, + 0.0041024405509233475, + 0.004430497530847788, + 0.061255697160959244, + 0.00597660057246685, + 0.004525675438344479, + 0.06545957177877426, + 0.007866699248552322, + 0.004777770955115557, + 0.004311115946620703, + 0.0014230962842702866, + 0.0030043944716453552, + 0.00251305871643126, + 0.002079624217003584, + 0.002119725104421377, + 0.0026568572502583265, + 0.0023392620496451855, + 0.0026736247818917036, + 0.0014504414284601808, + 0.005997942294925451, + 0.0028822161257267, + 0.0020459021907299757, + 0.008866067044436932, + 0.003470385679975152, + 0.003199776401743293, + 0.0021648225374519825, + 0.002201367635279894, + 0.0045115090906620026, + 0.0036268585827201605, + 0.004465010017156601, + 0.0025790445506572723, + 0.0030686447862535715, + 0.0034138141199946404, + 0.002761977259069681, + 0.002008303999900818, + 0.005729550030082464, + 0.0024507143534719944, + 0.003273103618994355, + 0.003286093706265092, + 0.00546661950647831, + 0.0048665981739759445, + 0.0021301452070474625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -1.7015693187713623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_tensor([9770, 9770])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_tensor([9770, 9770])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bfeb2091263711a097601ffd5d2a594ef8031faa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_tensor([9770, 9770])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium-length close-up shot from a horizontal angle focuses on a kind, middle-aged mother, standing quietly beside a carved, dark brown wooden window. Her graying hair is pulled back in a graceful bun, a few strands of silver caressing her ears. Her face is etched with the gentle wrinkles of time. She wears a simple, light beige cotton and linen blouse with a Chinese-style button-down pattern, delicately textured. A soft, afternoon sun filters through the window, casting dappled, warm shadows on her cheeks and shoulders, illuminating the fine dust particles floating in the air. She leans slightly toward the window, a gentle smile perpetually on her face, her lips naturally raised, her eyes etched with warm crow's feet. Her eyes are filled with love and tranquility, as if lost in beautiful memories. Beyond the window lies a hazy Jiangnan courtyard, a few stalks of green bamboo and a corner of a rising eaves faintly visible in the soft light. The main colors of the whole picture are warm brown and beige, creating a peaceful, warm and nostalgic atmosphere.", + "seed": 8955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_9770_seed_8955_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015039277309551836, + "vlm_logprob": -0.015039277309551836, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.015039277309551836, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02433699183166027 + ], + "opposite_logprob_tokens": [ + -3.77433705329895 + ], + "logprob_reward": -0.02433699183166027, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02433699183166027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27696022391319275, + 0.08416368812322617, + 0.04595015197992325, + 0.035779550671577454, + 0.1211053729057312, + 0.045447979122400284, + 0.06840308010578156, + 0.15615764260292053, + 0.03220241516828537, + 0.13382986187934875 + ], + "credit_map": [ + 0.034481488168239594, + 0.05396433547139168, + 0.0373067632317543, + 0.02855841815471649, + 0.04450007900595665, + 0.014350803568959236, + 0.03853776305913925, + 0.0252605602145195, + 0.010123246349394321, + 0.01605238951742649, + 0.01187052670866251, + 0.009718909859657288, + 0.01278418768197298, + 0.004902425687760115, + 0.011216029524803162, + 0.0074959686025977135, + 0.005919122602790594, + 0.004459818359464407, + 0.0057363081723451614, + 0.004980992991477251, + 0.008697666227817535, + 0.0033828557934612036, + 0.006276101805269718, + 0.006497284397482872, + 0.004518294241279364, + 0.004758709575980902, + 0.004509706050157547, + 0.0033201423939317465, + 0.005303666926920414, + 0.0022879708558321, + 0.006300628650933504, + 0.00478043407201767, + 0.031822048127651215, + 0.02521038055419922, + 0.023360813036561012, + 0.010343735106289387, + 0.011353409849107265, + 0.00271151727065444, + 0.009482876397669315, + 0.00682060094550252, + 0.005188637878745794, + 0.004357198718935251, + 0.004847407341003418, + 0.00413919473066926, + 0.006196416914463043, + 0.003389677032828331, + 0.009375124238431454, + 0.00795432087033987, + 0.020404107868671417, + 0.008674714714288712, + 0.0191756971180439, + 0.006094895303249359, + 0.005327660124748945, + 0.0018113666446879506, + 0.003903691889718175, + 0.0030109528452157974, + 0.05687171593308449, + 0.008752547204494476, + 0.05888619273900986, + 0.008022824302315712, + 0.008126839064061642, + 0.003422544803470373, + 0.0071031758561730385, + 0.0049718148075044155, + 0.008133349940180779, + 0.001768600894138217, + 0.0075296624563634396, + 0.004412371665239334, + 0.0020920687820762396, + 0.002499427180737257, + 0.0022581745870411396, + 0.00350875873118639, + 0.03054068796336651, + 0.012797262519598007, + 0.023929545655846596, + 0.010263930074870586, + 0.016725409775972366, + 0.004630958661437035, + 0.0240376815199852, + 0.010904368944466114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.02433699183166027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087371855974197 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087371855974197, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087371855974197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16335241496562958, + 0.08234409987926483, + 0.04268275201320648, + 0.03936322033405304, + 0.1759374588727951, + 0.04967738315463066, + 0.03443160653114319, + 0.28769993782043457, + 0.03561456874012947, + 0.08889657258987427 + ], + "credit_map": [ + 0.01777540147304535, + 0.0234297476708889, + 0.019252564758062363, + 0.01995762437582016, + 0.02704477310180664, + 0.009295448660850525, + 0.029686540365219116, + 0.016910308972001076, + 0.009932760149240494, + 0.006972246337682009, + 0.012184644117951393, + 0.010442563332617283, + 0.010716195218265057, + 0.006327862851321697, + 0.01697641983628273, + 0.008791404776275158, + 0.005760184954851866, + 0.003178513376042247, + 0.004982426296919584, + 0.009465271607041359, + 0.005168288480490446, + 0.0035965684801340103, + 0.005557505879551172, + 0.004973991308361292, + 0.004458732903003693, + 0.0034682368859648705, + 0.0044549903832376, + 0.006949171889573336, + 0.006090131588280201, + 0.0028902925550937653, + 0.006620170548558235, + 0.004431490320712328, + 0.044864725321531296, + 0.030762776732444763, + 0.03266338258981705, + 0.016711793839931488, + 0.01666109822690487, + 0.004137023817747831, + 0.017788628116250038, + 0.012348040007054806, + 0.006259091664105654, + 0.004825779236853123, + 0.005881797056645155, + 0.006600014865398407, + 0.008058268576860428, + 0.0026857415214180946, + 0.009335914626717567, + 0.006030770484358072, + 0.005931345745921135, + 0.003743732813745737, + 0.007887749001383781, + 0.005279114469885826, + 0.003988681361079216, + 0.0015864474698901176, + 0.003181041218340397, + 0.002833494683727622, + 0.11177113652229309, + 0.011178577318787575, + 0.11747976392507553, + 0.013488342985510826, + 0.012800119817256927, + 0.004817844368517399, + 0.00909313652664423, + 0.0070710210129618645, + 0.007862391881644726, + 0.001945784199051559, + 0.007535499054938555, + 0.004239276517182589, + 0.003385696792975068, + 0.0022666745353490114, + 0.00397682748734951, + 0.004402419086545706, + 0.026617275550961494, + 0.0076700784265995026, + 0.022794058546423912, + 0.0074034989811480045, + 0.008908377960324287, + 0.0036137213464826345, + 0.007330200634896755, + 0.004559364169836044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.016087371855974197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024888513144105673 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.0024888513144105673, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024888513144105673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15206581354141235, + 0.061933793127536774, + 0.06504513323307037, + 0.05131419748067856, + 0.10018125176429749, + 0.058469485491514206, + 0.03089938871562481, + 0.17352259159088135, + 0.17069648206233978, + 0.13587187230587006 + ], + "credit_map": [ + 0.025009410455822945, + 0.008431224152445793, + 0.025376416742801666, + 0.01638801023364067, + 0.022458979859948158, + 0.016000710427761078, + 0.021256420761346817, + 0.017144644632935524, + 0.009913340210914612, + 0.005333293229341507, + 0.011761324480175972, + 0.008942688815295696, + 0.008022968657314777, + 0.004599677864462137, + 0.006266223732382059, + 0.007094276137650013, + 0.010035167448222637, + 0.004995838273316622, + 0.010651624761521816, + 0.011766808107495308, + 0.0072836377657949924, + 0.006354780867695808, + 0.006336584687232971, + 0.007620692253112793, + 0.007453964091837406, + 0.004805705975741148, + 0.008293378166854382, + 0.011216224171221256, + 0.005259046331048012, + 0.0030149819795042276, + 0.005946156103163958, + 0.005324741825461388, + 0.028651686385273933, + 0.012100025080144405, + 0.018592532724142075, + 0.009538487531244755, + 0.008783633820712566, + 0.004106022417545319, + 0.005817392375320196, + 0.012591472826898098, + 0.010769794695079327, + 0.010218307375907898, + 0.00797524955123663, + 0.00782942958176136, + 0.006866564508527517, + 0.0032224818132817745, + 0.00662914989516139, + 0.004958510398864746, + 0.006776574533432722, + 0.0033275249879807234, + 0.007110207807272673, + 0.0037505333311855793, + 0.0023660799488425255, + 0.0021138133015483618, + 0.0032123480923473835, + 0.002242306713014841, + 0.05925821512937546, + 0.01015441957861185, + 0.06234391778707504, + 0.011523175984621048, + 0.01047433353960514, + 0.0054321796633303165, + 0.007942580617964268, + 0.006393767893314362, + 0.03236032649874687, + 0.012813587673008442, + 0.029291119426488876, + 0.029390936717391014, + 0.018036726862192154, + 0.018597936257719994, + 0.013100806623697281, + 0.01710503175854683, + 0.025698622688651085, + 0.01463533379137516, + 0.02470756508409977, + 0.026049546897411346, + 0.011518306098878384, + 0.010561814531683922, + 0.012629440985620022, + 0.010071253404021263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0024888513144105673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22450023889541626, + 0.05605599284172058, + 0.03862472251057625, + 0.03018428198993206, + 0.09443769603967667, + 0.07029858976602554, + 0.0435200035572052, + 0.32409968972206116, + 0.026014171540737152, + 0.0922645553946495 + ], + "credit_map": [ + 0.04476643353700638, + 0.03762997314333916, + 0.03742613643407822, + 0.02941255085170269, + 0.022593166679143906, + 0.0178816057741642, + 0.01748778112232685, + 0.01730259321630001, + 0.007774874567985535, + 0.0042997924610972404, + 0.00874375831335783, + 0.01129528321325779, + 0.0066176410764455795, + 0.005863531492650509, + 0.004998395219445229, + 0.006462715100497007, + 0.004626700188964605, + 0.0035577407106757164, + 0.004800910595804453, + 0.006670394446700811, + 0.005948412697762251, + 0.003044255543500185, + 0.0044446480460464954, + 0.005531660746783018, + 0.005125405732542276, + 0.002921606646850705, + 0.004730712156742811, + 0.004632002208381891, + 0.0044913445599377155, + 0.0018970550736412406, + 0.003621733281761408, + 0.002764422446489334, + 0.02532382868230343, + 0.011770721524953842, + 0.01541031152009964, + 0.008745712228119373, + 0.008936252444982529, + 0.003763826098293066, + 0.007435917388647795, + 0.013051124289631844, + 0.01073479000478983, + 0.006340102292597294, + 0.012511697597801685, + 0.007079567294567823, + 0.009263473562896252, + 0.0035383314825594425, + 0.012183801271021366, + 0.008646826259791851, + 0.009774468839168549, + 0.0032755215652287006, + 0.011970238760113716, + 0.0038561918772757053, + 0.004358928184956312, + 0.0025050928816199303, + 0.003988095559179783, + 0.003791461233049631, + 0.13034173846244812, + 0.00811934657394886, + 0.13758860528469086, + 0.014635586179792881, + 0.012096324935555458, + 0.004311022348701954, + 0.008126258850097656, + 0.008880793116986752, + 0.005143108777701855, + 0.002237503882497549, + 0.005190717987716198, + 0.0029678938444703817, + 0.0029986430890858173, + 0.0017703253543004394, + 0.0030067176558077335, + 0.0026992608327418566, + 0.03190067410469055, + 0.011618787422776222, + 0.026845552027225494, + 0.007749966345727444, + 0.004772295709699392, + 0.002096322597935796, + 0.004771374631673098, + 0.0025095788296312094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20364291965961456, + 0.03731336072087288, + 0.020967483520507812, + 0.021776804700493813, + 0.134372279047966, + 0.12102571874856949, + 0.075191430747509, + 0.14446358382701874, + 0.03255133703351021, + 0.20869505405426025 + ], + "credit_map": [ + 0.031031465157866478, + 0.032275017350912094, + 0.025809677317738533, + 0.025873150676488876, + 0.024476272985339165, + 0.018948141485452652, + 0.01829928532242775, + 0.026929935440421104, + 0.0036902823485434055, + 0.004032931290566921, + 0.004150166176259518, + 0.003747272305190563, + 0.006116137374192476, + 0.004083034582436085, + 0.00483675254508853, + 0.006656784564256668, + 0.0024382066912949085, + 0.003195498837158084, + 0.002521069021895528, + 0.002673909068107605, + 0.003521014703437686, + 0.0012358688982203603, + 0.002833377569913864, + 0.0025485395453870296, + 0.0033533594105392694, + 0.001553846406750381, + 0.002764767035841942, + 0.002184709534049034, + 0.003912797663360834, + 0.0020853658206760883, + 0.003466252703219652, + 0.0024557102005928755, + 0.033264074474573135, + 0.008125616237521172, + 0.019501741975545883, + 0.015522431582212448, + 0.012762853875756264, + 0.006357498001307249, + 0.01425088569521904, + 0.02458718977868557, + 0.021701866760849953, + 0.012984352186322212, + 0.024119723588228226, + 0.012570756487548351, + 0.013546661473810673, + 0.005687523167580366, + 0.01767432875931263, + 0.012740503996610641, + 0.012204533442854881, + 0.01433492824435234, + 0.011449798941612244, + 0.00872836448252201, + 0.009942140430212021, + 0.004060415085405111, + 0.007642962969839573, + 0.006828289479017258, + 0.058101754635572433, + 0.003218189114704728, + 0.0611056312918663, + 0.006247097626328468, + 0.005429981276392937, + 0.002754652639850974, + 0.0037059590686112642, + 0.003900330513715744, + 0.008166229352355003, + 0.0030365996062755585, + 0.0074380142614245415, + 0.003079627174884081, + 0.003689539385959506, + 0.0021121094468981028, + 0.0025414733681827784, + 0.0024877439718693495, + 0.07309690862894058, + 0.024805273860692978, + 0.06085529550909996, + 0.013685572892427444, + 0.012071439065039158, + 0.003433891572058201, + 0.014312870800495148, + 0.006433828733861446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.026334119960665703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length close-up shot from a horizontal angle focuses on a kind, middle-aged mother, standing quietly beside a carved, dark brown wooden window. Her graying hair is pulled back in a graceful bun, a few strands of silver caressing her ears. Her face is etched with the gentle wrinkles of time. She wears a simple, light beige cotton and linen blouse with a Chinese-style button-down pattern, delicately textured. A soft, afternoon sun filters through the window, casting dappled, warm shadows on her cheeks and shoulders, illuminating the fine dust particles floating in the air. She leans slightly toward the window, a gentle smile perpetually on her face, her lips naturally raised, her eyes etched with warm crow's feet. Her eyes are filled with love and tranquility, as if lost in beautiful memories. Beyond the window lies a hazy Jiangnan courtyard, a few stalks of green bamboo and a corner of a rising eaves faintly visible in the soft light. The main colors of the whole picture are warm brown and beige, creating a peaceful, warm and nostalgic atmosphere.", + "seed": 8956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_9770_seed_8956_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012748858146369457, + "vlm_logprob": -0.012748858146369457, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.012748858146369457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01937149278819561 + ], + "opposite_logprob_tokens": [ + -4.019371509552002 + ], + "logprob_reward": -0.01937149278819561, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01937149278819561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2694617509841919, + 0.08878116309642792, + 0.19289855659008026, + 0.03298627957701683, + 0.08360689133405685, + 0.10617022961378098, + 0.03410377725958824, + 0.04560502991080284, + 0.04614955559372902, + 0.10023680329322815 + ], + "credit_map": [ + 0.028272094205021858, + 0.02253192476928234, + 0.026424946263432503, + 0.030396824702620506, + 0.05743948370218277, + 0.030514944344758987, + 0.0338573195040226, + 0.040024206042289734, + 0.009216004982590675, + 0.008039409294724464, + 0.007123982533812523, + 0.010333136655390263, + 0.012528267689049244, + 0.00821473728865385, + 0.020291628316044807, + 0.01303398422896862, + 0.021634375676512718, + 0.024663738906383514, + 0.01666634902358055, + 0.02051701582968235, + 0.03248070552945137, + 0.014021811075508595, + 0.03545551747083664, + 0.02745901234447956, + 0.004127149004489183, + 0.003809681860730052, + 0.0029640118591487408, + 0.004040546249598265, + 0.004904784727841616, + 0.003311652923002839, + 0.005587826948612928, + 0.0042406232096254826, + 0.0076227555982768536, + 0.007605645339936018, + 0.006951786112040281, + 0.007130215875804424, + 0.01759067177772522, + 0.0084048081189394, + 0.01578066311776638, + 0.012520335614681244, + 0.008644758723676205, + 0.010310007259249687, + 0.008946439251303673, + 0.008343156427145004, + 0.02346912771463394, + 0.013048410415649414, + 0.018854107707738876, + 0.014554206281900406, + 0.0028993673622608185, + 0.00244498485699296, + 0.0035517741926014423, + 0.004558152053505182, + 0.00492698373273015, + 0.0028203155379742384, + 0.008341807872056961, + 0.004560388159006834, + 0.005594244226813316, + 0.006332123186439276, + 0.004005335737019777, + 0.005612727254629135, + 0.0057501341216266155, + 0.004395744763314724, + 0.007862353697419167, + 0.006052365526556969, + 0.0053694783709943295, + 0.00486141350120306, + 0.00660948408767581, + 0.005309485364705324, + 0.0067044212482869625, + 0.006998618599027395, + 0.0056676375679671764, + 0.004629014991223812, + 0.013154375366866589, + 0.009945318102836609, + 0.022404691204428673, + 0.017040561884641647, + 0.008653135038912296, + 0.009904857724905014, + 0.0095842769369483, + 0.009549569338560104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01937149278819561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557013891637325 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557013891637325, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557013891637325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27573955059051514, + 0.14995384216308594, + 0.08584222197532654, + 0.04807618632912636, + 0.0976390466094017, + 0.0789007619023323, + 0.03990887850522995, + 0.06705443561077118, + 0.06123533844947815, + 0.09564977884292603 + ], + "credit_map": [ + 0.03414158150553703, + 0.03081519342958927, + 0.03314802050590515, + 0.028917059302330017, + 0.032844677567481995, + 0.04635126143693924, + 0.03889128193259239, + 0.03063044510781765, + 0.013885597698390484, + 0.01420600526034832, + 0.014176388271152973, + 0.023324821144342422, + 0.013845788314938545, + 0.018537336960434914, + 0.027192216366529465, + 0.024785690009593964, + 0.005640419665724039, + 0.009107645601034164, + 0.006962975487112999, + 0.010002651251852512, + 0.01602936163544655, + 0.009341670200228691, + 0.014232617802917957, + 0.014524873346090317, + 0.004467890132218599, + 0.00641027744859457, + 0.005239749327301979, + 0.007379438728094101, + 0.007595003116875887, + 0.004911384079605341, + 0.006271845661103725, + 0.005800597369670868, + 0.010276096872985363, + 0.008898685686290264, + 0.015876006335020065, + 0.009941233322024345, + 0.018993571400642395, + 0.010019496083259583, + 0.013414754532277584, + 0.010219200514256954, + 0.00853400956839323, + 0.009817508980631828, + 0.011549944058060646, + 0.007244403474032879, + 0.012487618252635002, + 0.00965066161006689, + 0.010969830676913261, + 0.008646788075566292, + 0.004577559884637594, + 0.006642610300332308, + 0.0036801337264478207, + 0.005975564941763878, + 0.006531351245939732, + 0.0041370210237801075, + 0.004317200742661953, + 0.00404743105173111, + 0.007627812679857016, + 0.007657533045858145, + 0.005857457406818867, + 0.006281656678766012, + 0.008863161318004131, + 0.006739388685673475, + 0.010946924798190594, + 0.013080495409667492, + 0.008481099270284176, + 0.009643113240599632, + 0.007172050885856152, + 0.009181030094623566, + 0.006260404828935862, + 0.008051621727645397, + 0.0053845783695578575, + 0.00706144142895937, + 0.01257715467363596, + 0.012754792347550392, + 0.02135119028389454, + 0.017940063029527664, + 0.009288203902542591, + 0.00877224002033472, + 0.0064257183112204075, + 0.006540421396493912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012557013891637325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31564199924468994, + 0.1339612752199173, + 0.09873047471046448, + 0.05718525871634483, + 0.097189761698246, + 0.08327943086624146, + 0.05729464441537857, + 0.04858940467238426, + 0.04587502405047417, + 0.06225269287824631 + ], + "credit_map": [ + 0.03528305143117905, + 0.04921489581465721, + 0.0431143194437027, + 0.039449192583560944, + 0.040961746126413345, + 0.04506195709109306, + 0.032041046768426895, + 0.030515793710947037, + 0.019248096272349358, + 0.02563176117837429, + 0.015816019847989082, + 0.013857102021574974, + 0.010949891991913319, + 0.018252059817314148, + 0.01454420667141676, + 0.015662146732211113, + 0.020334476605057716, + 0.01527196541428566, + 0.021451562643051147, + 0.012953036464750767, + 0.007787197828292847, + 0.0055771032348275185, + 0.008701321668922901, + 0.0066538164392113686, + 0.007400827948004007, + 0.007554047275334597, + 0.009014340117573738, + 0.0055693332105875015, + 0.007875004783272743, + 0.005270077846944332, + 0.006906685885041952, + 0.0075949435122311115, + 0.015469069592654705, + 0.014510147273540497, + 0.01776544377207756, + 0.013711569830775261, + 0.010124881751835346, + 0.006193357985466719, + 0.007979391142725945, + 0.01143590547144413, + 0.011078950949013233, + 0.013920922763645649, + 0.012475523166358471, + 0.011327591724693775, + 0.006061811000108719, + 0.011270657181739807, + 0.008167129009962082, + 0.008976852521300316, + 0.00641323858872056, + 0.005743911024183035, + 0.008710391819477081, + 0.009737770073115826, + 0.008054851554334164, + 0.0052238404750823975, + 0.007042504847049713, + 0.006368138827383518, + 0.003361728973686695, + 0.007938847877085209, + 0.003621106268838048, + 0.008168967440724373, + 0.004109448287636042, + 0.005028075538575649, + 0.0059849051758646965, + 0.010376332327723503, + 0.005497874692082405, + 0.007279619574546814, + 0.005235543940216303, + 0.008974012918770313, + 0.006342178210616112, + 0.00392894446849823, + 0.0052472553215920925, + 0.0033695988822728395, + 0.008541627787053585, + 0.009657192975282669, + 0.0121138421818614, + 0.012091751210391521, + 0.0055121988989412785, + 0.004922708962112665, + 0.005277295596897602, + 0.004136077128350735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036149173974990845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3370177745819092, + 0.09168034046888351, + 0.04898590222001076, + 0.08507948368787766, + 0.12114608287811279, + 0.05546821281313896, + 0.06039731949567795, + 0.05634297430515289, + 0.08508533984422684, + 0.058796629309654236 + ], + "credit_map": [ + 0.050295911729335785, + 0.04105869308114052, + 0.033662352710962296, + 0.04983562231063843, + 0.050882916897535324, + 0.036247916519641876, + 0.03510665148496628, + 0.03992769867181778, + 0.010867975652217865, + 0.010245958343148232, + 0.015378695912659168, + 0.015236049890518188, + 0.008289974182844162, + 0.010154439136385918, + 0.011247840709984303, + 0.010259409435093403, + 0.006502706091850996, + 0.005831743124872446, + 0.007300202269107103, + 0.0072808885015547276, + 0.006053052842617035, + 0.004328723531216383, + 0.006615803111344576, + 0.005072782281786203, + 0.010575255379080772, + 0.012288855388760567, + 0.008458090014755726, + 0.009396943263709545, + 0.013595611788332462, + 0.00874033011496067, + 0.009232347831130028, + 0.012792048044502735, + 0.014776202850043774, + 0.015031295828521252, + 0.019256431609392166, + 0.015599869191646576, + 0.01882740668952465, + 0.012764733284711838, + 0.011518274433910847, + 0.013371866196393967, + 0.004945205990225077, + 0.010016190819442272, + 0.0054786112159490585, + 0.006186901591718197, + 0.006477693095803261, + 0.010618233121931553, + 0.005774575751274824, + 0.005970800295472145, + 0.010641108267009258, + 0.008158186450600624, + 0.006186747923493385, + 0.006385216023772955, + 0.005840476136654615, + 0.00725137535482645, + 0.008648731745779514, + 0.007285478990525007, + 0.006306502968072891, + 0.005811374168843031, + 0.0045367949642241, + 0.0061405799351632595, + 0.00676233135163784, + 0.006138225551694632, + 0.006931378971785307, + 0.013715784065425396, + 0.009150161407887936, + 0.008829876780509949, + 0.009210139513015747, + 0.012302912771701813, + 0.015659522265195847, + 0.006018899846822023, + 0.012707537971436977, + 0.01120628323405981, + 0.00544165913015604, + 0.005905434023588896, + 0.008402040228247643, + 0.009125931188464165, + 0.006947374902665615, + 0.00843624584376812, + 0.00848538801074028, + 0.006052556447684765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057168073952198 + ], + "opposite_logprob_tokens": [ + -3.895571708679199 + ], + "logprob_reward": -0.02057168073952198, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057168073952198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43784016370773315, + 0.13795728981494904, + 0.05804308131337166, + 0.04787420853972435, + 0.07484608888626099, + 0.04224482551217079, + 0.04969629645347595, + 0.04264567419886589, + 0.06547083705663681, + 0.04338150843977928 + ], + "credit_map": [ + 0.0863008126616478, + 0.04730701446533203, + 0.04353358596563339, + 0.054742805659770966, + 0.06368092447519302, + 0.03928765654563904, + 0.05140940472483635, + 0.05157797783613205, + 0.026038803160190582, + 0.015121127478778362, + 0.01619461178779602, + 0.013859898783266544, + 0.022792482748627663, + 0.012289539910852909, + 0.016850437968969345, + 0.014810392633080482, + 0.005795321427285671, + 0.006527079734951258, + 0.00802445039153099, + 0.007554370444267988, + 0.009323450736701488, + 0.005502487998455763, + 0.007228870410472155, + 0.00808704923838377, + 0.005177397280931473, + 0.006546122953295708, + 0.004710898268967867, + 0.0060013835318386555, + 0.008820897899568081, + 0.005622058175504208, + 0.005505003035068512, + 0.005490449722856283, + 0.0075251199305057526, + 0.01091771386563778, + 0.009431220591068268, + 0.010812061838805676, + 0.01172693446278572, + 0.0076187169179320335, + 0.00793530698865652, + 0.008879022672772408, + 0.004529458470642567, + 0.007715266663581133, + 0.004002560395747423, + 0.005037619266659021, + 0.004738654941320419, + 0.007304451894015074, + 0.004729533102363348, + 0.004187284503132105, + 0.006590041797608137, + 0.008155014365911484, + 0.0046576946042478085, + 0.007674476131796837, + 0.007719576824456453, + 0.003824300831183791, + 0.005378286354243755, + 0.0056969053111970425, + 0.005762894172221422, + 0.005272654816508293, + 0.004529539030045271, + 0.005440619774162769, + 0.005233124364167452, + 0.006160560064017773, + 0.005256307777017355, + 0.004989973735064268, + 0.006831162143498659, + 0.007024313788861036, + 0.010485169477760792, + 0.008803125470876694, + 0.011941147036850452, + 0.006094249896705151, + 0.0086152832955122, + 0.005676383152604103, + 0.006088320631533861, + 0.00449651712551713, + 0.0050538815557956696, + 0.00522235594689846, + 0.005638181231915951, + 0.0075386823154985905, + 0.005289968103170395, + 0.004053601995110512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02057168073952198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length close-up shot from a horizontal angle focuses on a kind, middle-aged mother, standing quietly beside a carved, dark brown wooden window. Her graying hair is pulled back in a graceful bun, a few strands of silver caressing her ears. Her face is etched with the gentle wrinkles of time. She wears a simple, light beige cotton and linen blouse with a Chinese-style button-down pattern, delicately textured. A soft, afternoon sun filters through the window, casting dappled, warm shadows on her cheeks and shoulders, illuminating the fine dust particles floating in the air. She leans slightly toward the window, a gentle smile perpetually on her face, her lips naturally raised, her eyes etched with warm crow's feet. Her eyes are filled with love and tranquility, as if lost in beautiful memories. Beyond the window lies a hazy Jiangnan courtyard, a few stalks of green bamboo and a corner of a rising eaves faintly visible in the soft light. The main colors of the whole picture are warm brown and beige, creating a peaceful, warm and nostalgic atmosphere.", + "seed": 8957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_9770_seed_8957_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014566167071461678, + "vlm_logprob": -0.014566167071461678, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.014566167071461678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021937096491456032 + ], + "opposite_logprob_tokens": [ + -3.896937131881714 + ], + "logprob_reward": -0.021937096491456032, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021937096491456032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28230518102645874, + 0.16828808188438416, + 0.06775754690170288, + 0.04283062741160393, + 0.08695521950721741, + 0.06142328679561615, + 0.02923094481229782, + 0.09505213797092438, + 0.06206107512116432, + 0.10409590601921082 + ], + "credit_map": [ + 0.02672514319419861, + 0.033862706273794174, + 0.038987308740615845, + 0.028917482122778893, + 0.039466798305511475, + 0.022294657304883003, + 0.06049865856766701, + 0.03155244141817093, + 0.012986988760530949, + 0.01107518095523119, + 0.021730616688728333, + 0.016763754189014435, + 0.02479100227355957, + 0.014238392934203148, + 0.04430406913161278, + 0.022398080676794052, + 0.004204771481454372, + 0.005327065475285053, + 0.006602173671126366, + 0.00822256226092577, + 0.009641949087381363, + 0.00574996555224061, + 0.018699631094932556, + 0.009309431537985802, + 0.003410348203033209, + 0.004396818112581968, + 0.007624197751283646, + 0.004320213571190834, + 0.006262397393584251, + 0.004212378989905119, + 0.007859048433601856, + 0.004745225887745619, + 0.00705610541626811, + 0.007778222672641277, + 0.011012645438313484, + 0.010293829254806042, + 0.012047423049807549, + 0.006816639564931393, + 0.02125008963048458, + 0.010700259357690811, + 0.005549591965973377, + 0.008094547316432, + 0.007275564596056938, + 0.008210304193198681, + 0.006139139644801617, + 0.00576230650767684, + 0.013188970275223255, + 0.0072028604336082935, + 0.0030218835454434156, + 0.0037170397117733955, + 0.004625513684004545, + 0.0035449659917503595, + 0.003674674779176712, + 0.003089547622948885, + 0.004052203148603439, + 0.003505116794258356, + 0.006745076272636652, + 0.008196631446480751, + 0.009790632873773575, + 0.011512829922139645, + 0.011916455812752247, + 0.008462759666144848, + 0.025821242481470108, + 0.012606506235897541, + 0.0036574150435626507, + 0.003804487641900778, + 0.007088320329785347, + 0.0066395103931427, + 0.009002525359392166, + 0.004363528452813625, + 0.019046973437070847, + 0.008458311669528484, + 0.008305724710226059, + 0.0064141456969082355, + 0.01526897307485342, + 0.008670213632285595, + 0.015061695128679276, + 0.008537149988114834, + 0.028230158612132072, + 0.013607841916382313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.021937096491456032 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210481852293015 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210481852293015, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210481852293015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23818521201610565, + 0.22071897983551025, + 0.11216358095407486, + 0.027567939832806587, + 0.06691714376211166, + 0.06278000771999359, + 0.03346412628889084, + 0.06173032894730568, + 0.0735996887087822, + 0.10287296772003174 + ], + "credit_map": [ + 0.01829601638019085, + 0.01828230358660221, + 0.03227867931127548, + 0.02660731039941311, + 0.024976495653390884, + 0.026245249435305595, + 0.0534365177154541, + 0.038062624633312225, + 0.017820097506046295, + 0.01451101154088974, + 0.02439095824956894, + 0.029344554990530014, + 0.026651352643966675, + 0.017694596201181412, + 0.05722733587026596, + 0.03307906910777092, + 0.010738392360508442, + 0.009029008448123932, + 0.014012040570378304, + 0.021455122157931328, + 0.008823092095553875, + 0.01391567662358284, + 0.014952925965189934, + 0.019237322732806206, + 0.0028065999504178762, + 0.002292506629601121, + 0.0028162184171378613, + 0.003083545947447419, + 0.0023561911657452583, + 0.00393297802656889, + 0.005304527003318071, + 0.004975371994078159, + 0.004611775279045105, + 0.004689250607043505, + 0.007837459444999695, + 0.008002062328159809, + 0.006978791672736406, + 0.006029844284057617, + 0.018611859530210495, + 0.010156099684536457, + 0.004995252471417189, + 0.004121125675737858, + 0.005962061695754528, + 0.00762649392709136, + 0.006831316277384758, + 0.005763705354183912, + 0.018137676641345024, + 0.0093423742800951, + 0.0033198886085301638, + 0.0027717917691916227, + 0.0052589187398552895, + 0.0035469636786729097, + 0.0043506016954779625, + 0.003773251548409462, + 0.006488191895186901, + 0.003954519517719746, + 0.004214364103972912, + 0.005691716447472572, + 0.005877223331481218, + 0.007270245347172022, + 0.007087399251759052, + 0.007513862568885088, + 0.014226174913346767, + 0.00984934065490961, + 0.004166231490671635, + 0.004640551749616861, + 0.008996470831334591, + 0.0075705633498728275, + 0.010385660454630852, + 0.005509539507329464, + 0.02206970751285553, + 0.01026095449924469, + 0.008070254698395729, + 0.0063820588402450085, + 0.01603308878839016, + 0.008688313886523247, + 0.013456376269459724, + 0.00975256972014904, + 0.02636231668293476, + 0.01412799023091793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.018210481852293015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951066501438618 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005951066501438618, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951066501438618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1997746229171753, + 0.08647343516349792, + 0.08034639805555344, + 0.039780452847480774, + 0.13298818469047546, + 0.2104310840368271, + 0.04540577158331871, + 0.09174830466508865, + 0.06822451204061508, + 0.044827260076999664 + ], + "credit_map": [ + 0.018838750198483467, + 0.016024645417928696, + 0.025531509891152382, + 0.02411339059472084, + 0.02354387566447258, + 0.024067813530564308, + 0.03941120207309723, + 0.028243422508239746, + 0.008038774132728577, + 0.006722345016896725, + 0.008581922389566898, + 0.009235004894435406, + 0.009163109585642815, + 0.013711672276258469, + 0.017599578946828842, + 0.01342102326452732, + 0.005770311690866947, + 0.00628238869830966, + 0.007076132111251354, + 0.007549102883785963, + 0.010652526281774044, + 0.010964264161884785, + 0.017707237973809242, + 0.014344434253871441, + 0.003490256378427148, + 0.0024547437205910683, + 0.003746061585843563, + 0.005137697793543339, + 0.0030733481980860233, + 0.006260282825678587, + 0.007853596471250057, + 0.007764467503875494, + 0.00857178121805191, + 0.008931237272918224, + 0.014928438700735569, + 0.01658993400633335, + 0.014292762614786625, + 0.013521353714168072, + 0.0375228226184845, + 0.018629856407642365, + 0.014383962377905846, + 0.012758656404912472, + 0.022838544100522995, + 0.025904295966029167, + 0.025769991800189018, + 0.015349600464105606, + 0.06413862109184265, + 0.029287420213222504, + 0.004769607447087765, + 0.003906574100255966, + 0.00863728579133749, + 0.0047666532918810844, + 0.006056084297597408, + 0.005304671358317137, + 0.0069975946098566055, + 0.004967303015291691, + 0.005997232627123594, + 0.007669833954423666, + 0.006421978119760752, + 0.011935579590499401, + 0.010897388681769371, + 0.009789478033781052, + 0.02529163844883442, + 0.013745172880589962, + 0.0041709644719958305, + 0.004679618403315544, + 0.007364865392446518, + 0.006822420749813318, + 0.01001446321606636, + 0.005532901734113693, + 0.020432643592357635, + 0.009206630289554596, + 0.004686220549046993, + 0.004576019011437893, + 0.005635432433336973, + 0.0038900242652744055, + 0.005358159076422453, + 0.006658812519162893, + 0.006641916465014219, + 0.007380678318440914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.005951066501438618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37823617458343506, + 0.14116200804710388, + 0.030452771112322807, + 0.05508190020918846, + 0.08139457553625107, + 0.028490476310253143, + 0.07699969410896301, + 0.0967547819018364, + 0.07034767419099808, + 0.041079938411712646 + ], + "credit_map": [ + 0.03429911285638809, + 0.022362753748893738, + 0.04439040645956993, + 0.03676708787679672, + 0.06123754754662514, + 0.030424049124121666, + 0.10158832371234894, + 0.047166891396045685, + 0.011710768565535545, + 0.008067670278251171, + 0.012840698473155499, + 0.016828756779432297, + 0.022041328251361847, + 0.01073592808097601, + 0.040331389755010605, + 0.018605468794703484, + 0.003097481094300747, + 0.002313331002369523, + 0.003121816087514162, + 0.003464719047769904, + 0.004308247938752174, + 0.004499258939176798, + 0.005720158573240042, + 0.003927761223167181, + 0.0038299383595585823, + 0.003011897671967745, + 0.006431020330637693, + 0.006660213228315115, + 0.005347406025975943, + 0.005697580985724926, + 0.016123628243803978, + 0.007980216294527054, + 0.005639122799038887, + 0.005903907585889101, + 0.008268703706562519, + 0.00966411828994751, + 0.010146004147827625, + 0.007609992753714323, + 0.022848881781101227, + 0.011313846334815025, + 0.003120808396488428, + 0.0036623752675950527, + 0.003512789262458682, + 0.0031871465034782887, + 0.004776414949446917, + 0.0033657154999673367, + 0.003610125510022044, + 0.003255104646086693, + 0.007704113610088825, + 0.006567616481333971, + 0.014309145510196686, + 0.00793280266225338, + 0.009761539287865162, + 0.009862784296274185, + 0.01167229749262333, + 0.00918940082192421, + 0.008476648479700089, + 0.009147596545517445, + 0.010975154116749763, + 0.011375104077160358, + 0.010534177534282207, + 0.012376305647194386, + 0.020793374627828598, + 0.013076421804726124, + 0.004194579552859068, + 0.005391733255237341, + 0.005266149528324604, + 0.007220347411930561, + 0.010683559812605381, + 0.006306284107267857, + 0.020764485001564026, + 0.010520544834434986, + 0.003299701726064086, + 0.003886146005243063, + 0.004538018722087145, + 0.003830746514722705, + 0.005863022990524769, + 0.004570808261632919, + 0.009132305160164833, + 0.0059591904282569885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191304333508015 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191304333508015, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191304333508015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36906230449676514, + 0.1529763787984848, + 0.03493209183216095, + 0.04087478667497635, + 0.022754017263650894, + 0.044194355607032776, + 0.039882298558950424, + 0.06437420099973679, + 0.1574167162179947, + 0.07353288680315018 + ], + "credit_map": [ + 0.05102873221039772, + 0.042288679629564285, + 0.03639908507466316, + 0.043106235563755035, + 0.03890020400285721, + 0.053649261593818665, + 0.04010690748691559, + 0.06358318030834198, + 0.013043778017163277, + 0.01030261255800724, + 0.013482254929840565, + 0.020376069471240044, + 0.021467557176947594, + 0.011862406507134438, + 0.04283948615193367, + 0.01960219070315361, + 0.003703956725075841, + 0.0033789852168411016, + 0.005600934382528067, + 0.0050385189242661, + 0.00388826965354383, + 0.003364426549524069, + 0.006033940706402063, + 0.0039230533875525, + 0.004030491691082716, + 0.003731457283720374, + 0.0035953661426901817, + 0.004470623563975096, + 0.005671361926943064, + 0.004779499489814043, + 0.009412427432835102, + 0.005183551926165819, + 0.002390844514593482, + 0.0025138966739177704, + 0.003096845466643572, + 0.0034429291263222694, + 0.0026865163818001747, + 0.002667347900569439, + 0.003438783111050725, + 0.0025168510619550943, + 0.0036070432979613543, + 0.0031701717525720596, + 0.004111676011234522, + 0.004839422646909952, + 0.007327159866690636, + 0.003850378794595599, + 0.01202908530831337, + 0.005259416066110134, + 0.003976901527494192, + 0.0031213145703077316, + 0.008432853035628796, + 0.004896861035376787, + 0.004353192634880543, + 0.004563633818179369, + 0.005903104320168495, + 0.004634433425962925, + 0.0056914882734417915, + 0.00517480680719018, + 0.008962336927652359, + 0.008462782949209213, + 0.004852675832808018, + 0.008248464204370975, + 0.014245463535189629, + 0.00873617921024561, + 0.009991577826440334, + 0.010906106792390347, + 0.016453705728054047, + 0.016857793554663658, + 0.02321913093328476, + 0.01168370433151722, + 0.046534713357686996, + 0.021769972518086433, + 0.005662354175001383, + 0.006173326633870602, + 0.007325639016926289, + 0.006248444318771362, + 0.007919158786535263, + 0.011881331913173199, + 0.013946712017059326, + 0.014375919476151466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.014191304333508015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length close-up shot from a horizontal angle focuses on a kind, middle-aged mother, standing quietly beside a carved, dark brown wooden window. Her graying hair is pulled back in a graceful bun, a few strands of silver caressing her ears. Her face is etched with the gentle wrinkles of time. She wears a simple, light beige cotton and linen blouse with a Chinese-style button-down pattern, delicately textured. A soft, afternoon sun filters through the window, casting dappled, warm shadows on her cheeks and shoulders, illuminating the fine dust particles floating in the air. She leans slightly toward the window, a gentle smile perpetually on her face, her lips naturally raised, her eyes etched with warm crow's feet. Her eyes are filled with love and tranquility, as if lost in beautiful memories. Beyond the window lies a hazy Jiangnan courtyard, a few stalks of green bamboo and a corner of a rising eaves faintly visible in the soft light. The main colors of the whole picture are warm brown and beige, creating a peaceful, warm and nostalgic atmosphere.", + "seed": 8958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_15/index_9770_seed_8958_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020930810645222665, + "vlm_logprob": -0.020930810645222665, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.020930810645222665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044781431555747986 + ], + "opposite_logprob_tokens": [ + -3.169781446456909 + ], + "logprob_reward": -0.044781431555747986, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044781431555747986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31029507517814636, + 0.12908190488815308, + 0.11048775911331177, + 0.04925759881734848, + 0.046685487031936646, + 0.06015276163816452, + 0.056934382766485214, + 0.07312025874853134, + 0.08790736645460129, + 0.07607736438512802 + ], + "credit_map": [ + 0.045329831540584564, + 0.04733673855662346, + 0.04251193627715111, + 0.037352852523326874, + 0.0492623932659626, + 0.022059476003050804, + 0.0381665863096714, + 0.02827523648738861, + 0.018780745565891266, + 0.013840783387422562, + 0.020564237609505653, + 0.014207576401531696, + 0.017019079998135567, + 0.00948674138635397, + 0.020371142774820328, + 0.014811605215072632, + 0.010745669715106487, + 0.009099358692765236, + 0.009605584666132927, + 0.00891804788261652, + 0.011122765950858593, + 0.022227833047509193, + 0.014241009019315243, + 0.02452748268842697, + 0.007494346238672733, + 0.004630922339856625, + 0.007521095685660839, + 0.00589070376008749, + 0.0065762680023908615, + 0.004391049034893513, + 0.006544315256178379, + 0.006208900362253189, + 0.0040015350095927715, + 0.00790486205369234, + 0.0021604876965284348, + 0.009963463991880417, + 0.006723818369209766, + 0.003834156086668372, + 0.0077655441127717495, + 0.0043316190131008625, + 0.004389308393001556, + 0.0067247855477035046, + 0.004865225870162249, + 0.0091341407969594, + 0.00911647453904152, + 0.007343176752328873, + 0.007241914514452219, + 0.011337731964886189, + 0.004077302757650614, + 0.003961591981351376, + 0.00922499317675829, + 0.005119688343256712, + 0.0061530619859695435, + 0.007893258705735207, + 0.013436870649456978, + 0.00706761097535491, + 0.007248076144605875, + 0.006042122840881348, + 0.010930076241493225, + 0.0056501710787415504, + 0.007862105034291744, + 0.012468849308788776, + 0.009183244779706001, + 0.013735620304942131, + 0.0065415953285992146, + 0.009095480665564537, + 0.01663804240524769, + 0.00954530294984579, + 0.00881495326757431, + 0.008783948607742786, + 0.0074538132175803185, + 0.021034229546785355, + 0.006028126925230026, + 0.01118931919336319, + 0.01114906556904316, + 0.013866938650608063, + 0.009146662428975105, + 0.0035990478936582804, + 0.016058534383773804, + 0.005039668176323175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.044781431555747986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317772895097733 + ], + "opposite_logprob_tokens": [ + -3.773317813873291 + ], + "logprob_reward": -0.023317772895097733, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317772895097733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20999208092689514, + 0.1908843070268631, + 0.10512588918209076, + 0.09710492938756943, + 0.03857920691370964, + 0.0801452249288559, + 0.09353571385145187, + 0.07660174369812012, + 0.07458042353391647, + 0.03345050662755966 + ], + "credit_map": [ + 0.030621647834777832, + 0.03337044268846512, + 0.01738925464451313, + 0.023202963173389435, + 0.03444622829556465, + 0.019282294437289238, + 0.025110522285103798, + 0.02656872570514679, + 0.02332070656120777, + 0.01793174259364605, + 0.022210678085684776, + 0.023367566987872124, + 0.02551259659230709, + 0.017792092636227608, + 0.037313707172870636, + 0.02343520149588585, + 0.01056463923305273, + 0.010443937964737415, + 0.008356290869414806, + 0.013816538266837597, + 0.009113085456192493, + 0.018542004749178886, + 0.012201095931231976, + 0.022088292986154556, + 0.006869759410619736, + 0.0038650333881378174, + 0.0035908473655581474, + 0.004203171003609896, + 0.011236501857638359, + 0.013295741751790047, + 0.02804984524846077, + 0.02599402889609337, + 0.003931671846657991, + 0.004731529857963324, + 0.0023139326367527246, + 0.005888999905437231, + 0.00505048967897892, + 0.005495734047144651, + 0.00530315050855279, + 0.005863698199391365, + 0.005583192687481642, + 0.006266230251640081, + 0.005177383776754141, + 0.0070442212745547295, + 0.010938319377601147, + 0.017659809440374374, + 0.007722865790128708, + 0.01975320279598236, + 0.006601730361580849, + 0.008251412771642208, + 0.015050139278173447, + 0.0122773302718997, + 0.011541537009179592, + 0.009014934301376343, + 0.02056630328297615, + 0.01023232750594616, + 0.007506819907575846, + 0.005771724972873926, + 0.013123063370585442, + 0.005744291935116053, + 0.008038869127631187, + 0.0131747517734766, + 0.0077955941669642925, + 0.015446623787283897, + 0.007490724790841341, + 0.004414824768900871, + 0.022510312497615814, + 0.0047836340963840485, + 0.004221091512590647, + 0.006404433865100145, + 0.009925882332026958, + 0.014829518273472786, + 0.0039054041262716055, + 0.004387029446661472, + 0.004420761950314045, + 0.004330013878643513, + 0.0047091529704630375, + 0.0028814964462071657, + 0.005371798295527697, + 0.003444849280640483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023317772895097733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23194065690040588, + 0.09795185923576355, + 0.06993725150823593, + 0.08963173627853394, + 0.11395242065191269, + 0.08498256653547287, + 0.1282864660024643, + 0.031867388635873795, + 0.09719331562519073, + 0.05425635352730751 + ], + "credit_map": [ + 0.02279573678970337, + 0.02190452627837658, + 0.015988463535904884, + 0.019596118479967117, + 0.027714649215340614, + 0.023665258660912514, + 0.05098048597574234, + 0.04929541051387787, + 0.010824903845787048, + 0.006593815982341766, + 0.008537723682820797, + 0.007409210782498121, + 0.010520800948143005, + 0.017162496224045753, + 0.012324604205787182, + 0.024578306823968887, + 0.006311243865638971, + 0.004663572646677494, + 0.004783191252499819, + 0.004393299575895071, + 0.007324887439608574, + 0.013921858742833138, + 0.0095833670347929, + 0.018955830484628677, + 0.006039584521204233, + 0.004368994385004044, + 0.0032812568824738264, + 0.0035536468494683504, + 0.011161101050674915, + 0.01096557080745697, + 0.027486031875014305, + 0.0227755568921566, + 0.00969921238720417, + 0.007118005305528641, + 0.013332701288163662, + 0.007728600408881903, + 0.008848371915519238, + 0.020374992862343788, + 0.016452834010124207, + 0.030397698283195496, + 0.004884655587375164, + 0.008284207433462143, + 0.005812115501612425, + 0.00900261476635933, + 0.011950427666306496, + 0.015614581294357777, + 0.00978742353618145, + 0.01964653842151165, + 0.009906282648444176, + 0.010225052945315838, + 0.018010295927524567, + 0.013064621947705746, + 0.016517791897058487, + 0.018925407901406288, + 0.025400785729289055, + 0.016236215829849243, + 0.0033515221439301968, + 0.00405754754319787, + 0.0034707277081906796, + 0.003573516383767128, + 0.004592122510075569, + 0.0030415947549045086, + 0.006164430174976587, + 0.0036159276496618986, + 0.005841932725161314, + 0.006395147647708654, + 0.012269279919564724, + 0.005661792121827602, + 0.010599292814731598, + 0.012824931181967258, + 0.01350817084312439, + 0.030092759057879448, + 0.005960691720247269, + 0.006407097447663546, + 0.0060465969145298, + 0.006395647767931223, + 0.006528295576572418, + 0.006706501357257366, + 0.008212177082896233, + 0.007999345660209656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639458566904068 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639458566904068, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639458566904068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2680857479572296, + 0.0884188860654831, + 0.03653522953391075, + 0.04144507274031639, + 0.05036698654294014, + 0.09702584892511368, + 0.13013456761837006, + 0.06956253945827484, + 0.17301779985427856, + 0.04540737718343735 + ], + "credit_map": [ + 0.02472083829343319, + 0.025418980047106743, + 0.01601618155837059, + 0.018701372668147087, + 0.04006553813815117, + 0.028815390542149544, + 0.0610220767557621, + 0.05332538112998009, + 0.009421833790838718, + 0.006958057172596455, + 0.008995305746793747, + 0.0064598447643220425, + 0.015278419479727745, + 0.009094374254345894, + 0.02120518684387207, + 0.011005868203938007, + 0.0028987431433051825, + 0.003093492239713669, + 0.002674343064427376, + 0.0037931674160063267, + 0.005353877786546946, + 0.005458778236061335, + 0.007885822094976902, + 0.00537700904533267, + 0.004388097207993269, + 0.004016620106995106, + 0.0030353907495737076, + 0.0037607294507324696, + 0.005613008514046669, + 0.005869447253644466, + 0.00686688581481576, + 0.007894893176853657, + 0.0034822432789951563, + 0.005490264855325222, + 0.0026637353003025055, + 0.0054268548265099525, + 0.005731556564569473, + 0.008972890675067902, + 0.007210691925138235, + 0.011388747952878475, + 0.006622912362217903, + 0.009744341485202312, + 0.004098590929061174, + 0.009088885970413685, + 0.016407402232289314, + 0.01589849777519703, + 0.011750232428312302, + 0.023414986208081245, + 0.012331448495388031, + 0.009407814592123032, + 0.023006340488791466, + 0.013474603183567524, + 0.01681215688586235, + 0.013632725924253464, + 0.027757899835705757, + 0.013711572624742985, + 0.005824646446853876, + 0.004086559172719717, + 0.007993698120117188, + 0.003539649536833167, + 0.009662872180342674, + 0.01211255881935358, + 0.01193283498287201, + 0.014409719035029411, + 0.01853218674659729, + 0.004762392025440931, + 0.05582438409328461, + 0.006472582928836346, + 0.008738725446164608, + 0.014967508614063263, + 0.01865694299340248, + 0.0450630709528923, + 0.0036787488497793674, + 0.00453883595764637, + 0.007737310137599707, + 0.006111128255724907, + 0.006873826961964369, + 0.0035080884117633104, + 0.008462038822472095, + 0.004497396759688854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.008639458566904068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged woman with graying hair styled in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a light beige cotton and linen blouse with a Chinese-style button-down pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set beside a carved, dark brown wooden window with a hazy Jiangnan courtyard visible outside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft afternoon sun casting dappled shadows and illuminating floating dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-length close-up from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327747642993927 + ], + "opposite_logprob_tokens": [ + -3.773277521133423 + ], + "logprob_reward": -0.02327747642993927, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327747642993927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2742611765861511, + 0.09116198867559433, + 0.06730058789253235, + 0.15167109668254852, + 0.051513269543647766, + 0.07317840307950974, + 0.11466659605503082, + 0.05616888403892517, + 0.07368703931570053, + 0.04639101028442383 + ], + "credit_map": [ + 0.030196944251656532, + 0.036319002509117126, + 0.026445139199495316, + 0.036999642848968506, + 0.03425973281264305, + 0.02987208403646946, + 0.041083093732595444, + 0.03908553346991539, + 0.009080243296921253, + 0.00908596906810999, + 0.007391059305518866, + 0.008453755639493465, + 0.015571017749607563, + 0.009034808725118637, + 0.018241098150610924, + 0.01430402509868145, + 0.0070478422567248344, + 0.006933624856173992, + 0.004972260911017656, + 0.008636820130050182, + 0.00708337128162384, + 0.01205751858651638, + 0.007940747775137424, + 0.012628396973013878, + 0.013169532641768456, + 0.008026416413486004, + 0.008971729315817356, + 0.00741208903491497, + 0.02008921466767788, + 0.015193521045148373, + 0.048955418169498444, + 0.029853159561753273, + 0.004863538313657045, + 0.0063659739680588245, + 0.0029159453697502613, + 0.007582317106425762, + 0.009295476600527763, + 0.00535681564360857, + 0.008519291877746582, + 0.006613904610276222, + 0.006682350300252438, + 0.01004908699542284, + 0.007102319970726967, + 0.012270049192011356, + 0.011085011065006256, + 0.008473161607980728, + 0.009652761742472649, + 0.00786365196108818, + 0.009190450422465801, + 0.00757154356688261, + 0.008807103149592876, + 0.009035331197082996, + 0.01581314019858837, + 0.02469010278582573, + 0.015569412149488926, + 0.02398950792849064, + 0.006308244541287422, + 0.0058432226069271564, + 0.009577495977282524, + 0.006216761656105518, + 0.0058138021267950535, + 0.006159530021250248, + 0.008102414198219776, + 0.008147402666509151, + 0.004680962767452002, + 0.006169903092086315, + 0.00490222591906786, + 0.0068953391164541245, + 0.00817861594259739, + 0.008711138740181923, + 0.017288966104388237, + 0.016859887167811394, + 0.0038788123056292534, + 0.005535597447305918, + 0.007859152741730213, + 0.004521029070019722, + 0.008671524003148079, + 0.003052959218621254, + 0.007065560203045607, + 0.005806367844343185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.02327747642993927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_tensor([1691, 1691])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_tensor([1691, 1691])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4875ba3364ad58fbc22ad92b7f52d9057cd2410d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_tensor([1691, 1691])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a medium shot, shot from a slightly upward angle, capturing an elderly woman with silver hair tied in a simple bun, nestled in a tranquil Chinese courtyard in late autumn. She wears a dark blue, double-breasted, coarse cloth training suit, the fabric appearing worn yet crisp and neat. She exudes a timeless calmness. Golden ginkgo leaves and crimson maple leaves slowly fall from sparse branches, forming a thick, vibrant carpet beneath her feet. Her hands steady a pristine longsword, its hilt clad in dark wood, aged by time. The brass guard gleams with a dull luster in the afternoon sun. She lowers herself, her center of gravity steady, her gaze fixed firmly ahead. With a slow, restrained, and powerful turn, the ancient sword traces a steady, flowing arc through the air, its wind gently sweeping up a few fallen leaves, swirling with them. At this point, the camera begins to move in very slowly and smoothly, gradually blurring the background courtyard and focusing entirely on her. Finally, the camera lingers on a close-up of her weathered face, each deep wrinkle telling a long story. Her slightly cloudy yet eagle-like eyes gleam with unyielding focus. The entire scene, bathed in soft, warm light, creates a tranquil atmosphere of solemnity, focus, and power.", + "seed": 14139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14139_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.055812244582921264, + "vlm_logprob": -0.055812244582921264, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.055812244582921264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0817314088344574 + ], + "opposite_logprob_tokens": [ + -2.5817313194274902 + ], + "logprob_reward": -0.0817314088344574, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0817314088344574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08090721070766449, + 0.11964191496372223, + 0.036219522356987, + 0.06779761612415314, + 0.07441996037960052, + 0.06749987602233887, + 0.12190359085798264, + 0.1873413473367691, + 0.07410872727632523, + 0.17016024887561798 + ], + "credit_map": [ + 0.010494114831089973, + 0.01042621023952961, + 0.009377039037644863, + 0.011311857961118221, + 0.014031633734703064, + 0.00886478740721941, + 0.009458515793085098, + 0.006943048909306526, + 0.010724131017923355, + 0.014809892512857914, + 0.02511598914861679, + 0.026880137622356415, + 0.010838613845407963, + 0.010545185767114162, + 0.011085244826972485, + 0.009642720222473145, + 0.005668337922543287, + 0.004630402196198702, + 0.0048009371384978294, + 0.004722722806036472, + 0.004835854284465313, + 0.0040441337041556835, + 0.0036204224452376366, + 0.0038967139553278685, + 0.008120517246425152, + 0.007041845936328173, + 0.005626391619443893, + 0.008432717993855476, + 0.012699929066002369, + 0.008319903165102005, + 0.008514359593391418, + 0.00904196035116911, + 0.011280748061835766, + 0.005583972670137882, + 0.007761180866509676, + 0.006562039256095886, + 0.013096891343593597, + 0.011543963104486465, + 0.011035040952265263, + 0.007556121796369553, + 0.012053363025188446, + 0.0045238365419209, + 0.016350382938981056, + 0.004745211917907, + 0.009731595404446125, + 0.003498960519209504, + 0.012203464284539223, + 0.004393064416944981, + 0.022590631619095802, + 0.014339101500809193, + 0.021447891369462013, + 0.008828980848193169, + 0.012886587530374527, + 0.021162303164601326, + 0.012955068610608578, + 0.007693026214838028, + 0.022817067801952362, + 0.017939720302820206, + 0.02416626177728176, + 0.020756365731358528, + 0.027845583856105804, + 0.02497025579214096, + 0.02079148218035698, + 0.0280546173453331, + 0.006029338110238314, + 0.015603840351104736, + 0.005248828791081905, + 0.009178257547318935, + 0.013908836990594864, + 0.009149780496954918, + 0.009891911409795284, + 0.005097934044897556, + 0.009794101119041443, + 0.030924178659915924, + 0.010277283377945423, + 0.02380543388426304, + 0.03053463064134121, + 0.023050641641020775, + 0.029203685000538826, + 0.012570290826261044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0817314088344574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029831413179636 + ], + "opposite_logprob_tokens": [ + -3.5298314094543457 + ], + "logprob_reward": -0.029831413179636, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029831413179636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1030111089348793, + 0.057030774652957916, + 0.04961991682648659, + 0.06911993771791458, + 0.059595778584480286, + 0.24244141578674316, + 0.23249444365501404, + 0.03681090474128723, + 0.045650020241737366, + 0.10422573983669281 + ], + "credit_map": [ + 0.012168793007731438, + 0.016458211466670036, + 0.013904744759202003, + 0.01717466674745083, + 0.011744612827897072, + 0.011275392025709152, + 0.012427610345184803, + 0.007857071235775948, + 0.006084315944463015, + 0.007037293165922165, + 0.010939904488623142, + 0.013643846847116947, + 0.003772233845666051, + 0.004903044551610947, + 0.004834238905459642, + 0.00581589387729764, + 0.009171963669359684, + 0.006935436278581619, + 0.006138390861451626, + 0.006604074966162443, + 0.005530935246497393, + 0.004979465622454882, + 0.005210822448134422, + 0.005048821214586496, + 0.008272497914731503, + 0.005688606761395931, + 0.005253375042229891, + 0.01042388379573822, + 0.014109959825873375, + 0.008230983279645443, + 0.008646213449537754, + 0.00849441159516573, + 0.010898487642407417, + 0.00575082516297698, + 0.005698644556105137, + 0.004482165444642305, + 0.010928979143500328, + 0.007597756572067738, + 0.009030316956341267, + 0.005208600778132677, + 0.042283717542886734, + 0.014143922366201878, + 0.07397706806659698, + 0.016706837341189384, + 0.022550566121935844, + 0.011411682702600956, + 0.04429172724485397, + 0.017075859010219574, + 0.04782450199127197, + 0.02586429938673973, + 0.03819356858730316, + 0.016657287254929543, + 0.020185576751828194, + 0.04148225486278534, + 0.02597225457429886, + 0.0163146760314703, + 0.004466468933969736, + 0.00399828189983964, + 0.004078611731529236, + 0.00469923485070467, + 0.005933910142630339, + 0.004854080267250538, + 0.004375568125396967, + 0.004404745530337095, + 0.00461106700822711, + 0.0061574773862957954, + 0.004570308607071638, + 0.00617254851385951, + 0.008220191113650799, + 0.004756958689540625, + 0.0081187067553401, + 0.0030427558813244104, + 0.008638815954327583, + 0.015923533588647842, + 0.01040472649037838, + 0.010804749093949795, + 0.016342245042324066, + 0.014201527461409569, + 0.020407475531101227, + 0.00750264897942543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.029831413179636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23560938239097595, + 0.12217359244823456, + 0.05406704172492027, + 0.11905362457036972, + 0.04110613837838173, + 0.1397639364004135, + 0.11837837100028992, + 0.03515506163239479, + 0.06522896140813828, + 0.06946395337581635 + ], + "credit_map": [ + 0.02887299470603466, + 0.039758261293172836, + 0.03298952803015709, + 0.04875371977686882, + 0.02241547405719757, + 0.026523351669311523, + 0.016665171831846237, + 0.01963087171316147, + 0.010765112936496735, + 0.010418053716421127, + 0.028554532676935196, + 0.03253871947526932, + 0.00862370990216732, + 0.010408507660031319, + 0.010332036763429642, + 0.010532917454838753, + 0.0117418821901083, + 0.007295865565538406, + 0.007473119534552097, + 0.006442977115511894, + 0.004525011405348778, + 0.005580289289355278, + 0.004100719932466745, + 0.006907179951667786, + 0.021765295416116714, + 0.015538896434009075, + 0.016483796760439873, + 0.014331082813441753, + 0.010196010582149029, + 0.01814710535109043, + 0.011199111118912697, + 0.011392318643629551, + 0.009122619405388832, + 0.0048204343765974045, + 0.003829715773463249, + 0.004166817292571068, + 0.0036848713643848896, + 0.0052293394692242146, + 0.005764758680015802, + 0.004487581551074982, + 0.024885062128305435, + 0.007609435357153416, + 0.04370851069688797, + 0.009330647997558117, + 0.012028870172798634, + 0.006340872962027788, + 0.026253998279571533, + 0.009606553241610527, + 0.0241115540266037, + 0.013846180401742458, + 0.01968037150800228, + 0.00832976121455431, + 0.011822222732007504, + 0.019914833828806877, + 0.012539424933493137, + 0.008134028874337673, + 0.003973869141191244, + 0.004690106958150864, + 0.0032768642995506525, + 0.0032917640637606382, + 0.006345098372548819, + 0.003869952168315649, + 0.004202425014227629, + 0.005504980217665434, + 0.010186657309532166, + 0.007744709961116314, + 0.016383664682507515, + 0.006504006218165159, + 0.00778920017182827, + 0.005870601627975702, + 0.006516897585242987, + 0.004233218729496002, + 0.006015725899487734, + 0.013051985763013363, + 0.006172642577439547, + 0.012044758535921574, + 0.008594519458711147, + 0.0099748270586133, + 0.008367669768631458, + 0.005241820123046637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281100034713745 + ], + "opposite_logprob_tokens": [ + -2.0178110599517822 + ], + "logprob_reward": -0.14281100034713745, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281100034713745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496548056602478, + 0.0907774418592453, + 0.1505947709083557, + 0.11512327939271927, + 0.10243580490350723, + 0.1267329752445221, + 0.09853093326091766, + 0.03740148991346359, + 0.03313802555203438, + 0.09561048448085785 + ], + "credit_map": [ + 0.01738087832927704, + 0.024591144174337387, + 0.01339792925864458, + 0.01602170430123806, + 0.02751106023788452, + 0.01842794381082058, + 0.01691083423793316, + 0.015413304790854454, + 0.011701175011694431, + 0.016245849430561066, + 0.010670328512787819, + 0.014058038592338562, + 0.009277179837226868, + 0.00940382108092308, + 0.010986298322677612, + 0.008434747345745564, + 0.023339178413152695, + 0.023545952513813972, + 0.021829117089509964, + 0.016569076105952263, + 0.016841834411025047, + 0.014801323413848877, + 0.018371976912021637, + 0.015296312049031258, + 0.014527466148138046, + 0.011596942320466042, + 0.011786560527980328, + 0.018525172024965286, + 0.01903865486383438, + 0.010991842485964298, + 0.015357820317149162, + 0.013298817910254002, + 0.024439748376607895, + 0.0077053760178387165, + 0.012307568453252316, + 0.010165371932089329, + 0.009110153652727604, + 0.013233275152742863, + 0.014272061176598072, + 0.0112022515386343, + 0.020808467641472816, + 0.006821208633482456, + 0.04015360772609711, + 0.0085349902510643, + 0.009786644019186497, + 0.007764918264001608, + 0.023670874536037445, + 0.009192266501486301, + 0.01800057664513588, + 0.011001266539096832, + 0.014225081540644169, + 0.008493096567690372, + 0.013172969222068787, + 0.014981275424361229, + 0.01166600827127695, + 0.006990658584982157, + 0.006910930387675762, + 0.003906660713255405, + 0.005056031979620457, + 0.00425295252352953, + 0.004723730497062206, + 0.0036929529160261154, + 0.003685466945171356, + 0.00517276581376791, + 0.0036013515200465918, + 0.004000930115580559, + 0.0057578012347221375, + 0.0044702282175421715, + 0.003946962300688028, + 0.0031889337114989758, + 0.00479627912864089, + 0.0033755400218069553, + 0.0076934657990932465, + 0.012931517325341702, + 0.008836977183818817, + 0.017718177288770676, + 0.012558362446725368, + 0.010867606848478317, + 0.01853509433567524, + 0.006469282321631908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.14281100034713745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08826883137226105, + 0.12862883508205414, + 0.09534641355276108, + 0.10207168757915497, + 0.08580808341503143, + 0.11409134417772293, + 0.0420069582760334, + 0.09566015005111694, + 0.04276641830801964, + 0.20535123348236084 + ], + "credit_map": [ + 0.012001375667750835, + 0.013743504881858826, + 0.013383962213993073, + 0.010430391877889633, + 0.010191299021244049, + 0.009947025217115879, + 0.010597850196063519, + 0.007973420433700085, + 0.014885775744915009, + 0.017261359840631485, + 0.02067485824227333, + 0.023268356919288635, + 0.013580631464719772, + 0.011472736485302448, + 0.014325140975415707, + 0.013159966096282005, + 0.01313075516372919, + 0.010526204481720924, + 0.013289877213537693, + 0.008674688637256622, + 0.017352087423205376, + 0.00895787961781025, + 0.013337350450456142, + 0.01007755845785141, + 0.012993000447750092, + 0.012030867859721184, + 0.010435044765472412, + 0.014487415552139282, + 0.018019715324044228, + 0.009343497455120087, + 0.012204482220113277, + 0.01255766674876213, + 0.017101779580116272, + 0.0065041352063417435, + 0.008971912786364555, + 0.010783915407955647, + 0.011956405825912952, + 0.008391107432544231, + 0.013682076707482338, + 0.00841674953699112, + 0.016064323484897614, + 0.007734420709311962, + 0.03284401446580887, + 0.0083414101973176, + 0.01189939584583044, + 0.006605709902942181, + 0.020828677341341972, + 0.00977338571101427, + 0.005894619971513748, + 0.006049083545804024, + 0.0036620846949517727, + 0.0037230055313557386, + 0.006175358314067125, + 0.008535627275705338, + 0.004792911931872368, + 0.003174266079440713, + 0.01353753637522459, + 0.010741314850747585, + 0.009758715517818928, + 0.010143381543457508, + 0.01471240445971489, + 0.012880288995802402, + 0.01098696794360876, + 0.012899540364742279, + 0.005266848485916853, + 0.005277675576508045, + 0.007407656870782375, + 0.004631428513675928, + 0.006648113485425711, + 0.0039667957462370396, + 0.005119600798934698, + 0.004448301158845425, + 0.013705608434975147, + 0.03343728557229042, + 0.018077025189995766, + 0.05030635744333267, + 0.023006973788142204, + 0.025844696909189224, + 0.02573644556105137, + 0.01523683499544859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.004096450749784708 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from a slightly upward angle, capturing an elderly woman with silver hair tied in a simple bun, nestled in a tranquil Chinese courtyard in late autumn. She wears a dark blue, double-breasted, coarse cloth training suit, the fabric appearing worn yet crisp and neat. She exudes a timeless calmness. Golden ginkgo leaves and crimson maple leaves slowly fall from sparse branches, forming a thick, vibrant carpet beneath her feet. Her hands steady a pristine longsword, its hilt clad in dark wood, aged by time. The brass guard gleams with a dull luster in the afternoon sun. She lowers herself, her center of gravity steady, her gaze fixed firmly ahead. With a slow, restrained, and powerful turn, the ancient sword traces a steady, flowing arc through the air, its wind gently sweeping up a few fallen leaves, swirling with them. At this point, the camera begins to move in very slowly and smoothly, gradually blurring the background courtyard and focusing entirely on her. Finally, the camera lingers on a close-up of her weathered face, each deep wrinkle telling a long story. Her slightly cloudy yet eagle-like eyes gleam with unyielding focus. The entire scene, bathed in soft, warm light, creates a tranquil atmosphere of solemnity, focus, and power.", + "seed": 14140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14140_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.054934443160891536, + "vlm_logprob": -0.054934443160891536, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.054934443160891536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04219670593738556 + ], + "opposite_logprob_tokens": [ + -3.292196750640869 + ], + "logprob_reward": -0.04219670593738556, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04219670593738556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13841597735881805, + 0.02564409375190735, + 0.10629528760910034, + 0.16890376806259155, + 0.12776216864585876, + 0.13772499561309814, + 0.08762587606906891, + 0.10501272976398468, + 0.05947328731417656, + 0.04314178600907326 + ], + "credit_map": [ + 0.016316117718815804, + 0.017638014629483223, + 0.023848241195082664, + 0.018266990780830383, + 0.014291197061538696, + 0.013701134361326694, + 0.013447677716612816, + 0.0209065992385149, + 0.003091383259743452, + 0.0027816984802484512, + 0.004679608158767223, + 0.002902538748458028, + 0.002125457627698779, + 0.0029828492552042007, + 0.0030610552057623863, + 0.004019503481686115, + 0.02163066528737545, + 0.011214416474103928, + 0.011836604215204716, + 0.014557084068655968, + 0.008876358158886433, + 0.012896648608148098, + 0.011825875379145145, + 0.013457631692290306, + 0.03837655857205391, + 0.020576607435941696, + 0.02162519469857216, + 0.01745964214205742, + 0.012561588548123837, + 0.023448312655091286, + 0.013726351782679558, + 0.02112950012087822, + 0.012657892890274525, + 0.021427318453788757, + 0.012512614019215107, + 0.01863122545182705, + 0.013920780271291733, + 0.017741870135068893, + 0.0127715477719903, + 0.018098922446370125, + 0.012941772118210793, + 0.024071767926216125, + 0.023010168224573135, + 0.017839685082435608, + 0.014338775537908077, + 0.011184883303940296, + 0.014397071674466133, + 0.019940877333283424, + 0.006253387778997421, + 0.015486449934542179, + 0.009763806127011776, + 0.01278578583151102, + 0.017398588359355927, + 0.005568120628595352, + 0.009172244928777218, + 0.011197490617632866, + 0.007479786407202482, + 0.017355645075440407, + 0.009795782156288624, + 0.013552830554544926, + 0.029385825619101524, + 0.00822884775698185, + 0.009099612943828106, + 0.010114406235516071, + 0.004244472831487656, + 0.007750950753688812, + 0.0073463451117277145, + 0.00809695664793253, + 0.01648779585957527, + 0.005266753025352955, + 0.0060754576697945595, + 0.004204559605568647, + 0.006513923406600952, + 0.004554602783173323, + 0.005659504793584347, + 0.005759906023740768, + 0.006687081418931484, + 0.004546452779322863, + 0.005023747682571411, + 0.0043965671211481094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.04219670593738556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212588503956795 + ], + "opposite_logprob_tokens": [ + -4.018212795257568 + ], + "logprob_reward": -0.018212588503956795, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018212588503956795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1705310046672821, + 0.05110178887844086, + 0.1455928236246109, + 0.07122442126274109, + 0.22588932514190674, + 0.2137845754623413, + 0.01434171199798584, + 0.047445591539144516, + 0.026249906048178673, + 0.033838868141174316 + ], + "credit_map": [ + 0.020246382802724838, + 0.01956333965063095, + 0.029689835384488106, + 0.02286323346197605, + 0.018176311627030373, + 0.016850577667355537, + 0.01685338094830513, + 0.026287958025932312, + 0.006755849812179804, + 0.00790379848331213, + 0.007875490002334118, + 0.005434392485767603, + 0.006045321933925152, + 0.005664830096065998, + 0.005363546311855316, + 0.006058562081307173, + 0.032481543719768524, + 0.0158625990152359, + 0.01472405344247818, + 0.01850138232111931, + 0.011393197812139988, + 0.019198233261704445, + 0.01485532894730568, + 0.018576502799987793, + 0.008290682919323444, + 0.01309176441282034, + 0.005225281231105328, + 0.007958406582474709, + 0.007978404872119427, + 0.011114824563264847, + 0.006807800382375717, + 0.010757267475128174, + 0.027200903743505478, + 0.021877657622098923, + 0.04465553164482117, + 0.032517753541469574, + 0.017993828281760216, + 0.019348163157701492, + 0.028711451217532158, + 0.03358406573534012, + 0.019242936745285988, + 0.03837055712938309, + 0.04088493809103966, + 0.024604883044958115, + 0.022095948457717896, + 0.020433353260159492, + 0.01960345171391964, + 0.02854852005839348, + 0.0013881339691579342, + 0.0023389689158648252, + 0.0022017406299710274, + 0.002026820555329323, + 0.0019139559008181095, + 0.0016871890984475613, + 0.0014833577442914248, + 0.0013015457661822438, + 0.00579878780990839, + 0.005160142667591572, + 0.004814016632735729, + 0.006187804974615574, + 0.007752974051982164, + 0.006033048965036869, + 0.006204791367053986, + 0.005494032520800829, + 0.0023650226648896933, + 0.002585766837000847, + 0.002608603797852993, + 0.004735283087939024, + 0.004269747529178858, + 0.003509379690513015, + 0.0036142258904874325, + 0.002561879577115178, + 0.003675640793517232, + 0.0033307953272014856, + 0.00351471989415586, + 0.004976471420377493, + 0.006635394878685474, + 0.0035143441054970026, + 0.0043399822898209095, + 0.0038515240885317326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.018212588503956795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298904299736023 + ], + "opposite_logprob_tokens": [ + -3.773298978805542 + ], + "logprob_reward": -0.023298904299736023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298904299736023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20726968348026276, + 0.061209358274936676, + 0.1156909167766571, + 0.07891495525836945, + 0.16532056033611298, + 0.18186278641223907, + 0.027940861880779266, + 0.029698997735977173, + 0.06837834417819977, + 0.06371352076530457 + ], + "credit_map": [ + 0.037704113870859146, + 0.02785315178334713, + 0.031645648181438446, + 0.017098387703299522, + 0.014096739701926708, + 0.024032844230532646, + 0.01570345088839531, + 0.03913533687591553, + 0.006024713162332773, + 0.00643262080848217, + 0.0149780148640275, + 0.004728467669337988, + 0.006497792899608612, + 0.0064548589289188385, + 0.005727272946387529, + 0.010365620255470276, + 0.02219167724251747, + 0.012104872614145279, + 0.013035977259278297, + 0.01573260873556137, + 0.010245570912957191, + 0.015592187643051147, + 0.01247638650238514, + 0.014311632141470909, + 0.013791746459901333, + 0.012113342061638832, + 0.006473845802247524, + 0.007053373847156763, + 0.007692353334277868, + 0.011611703783273697, + 0.0077323224395513535, + 0.01244626846164465, + 0.01966296136379242, + 0.016319621354341507, + 0.031752392649650574, + 0.023400338366627693, + 0.013102556578814983, + 0.015443836338818073, + 0.02014448679983616, + 0.025494372472167015, + 0.016324302181601524, + 0.031656742095947266, + 0.03582897037267685, + 0.02122710645198822, + 0.017349233850836754, + 0.01567908562719822, + 0.01770441234111786, + 0.026092929765582085, + 0.002949761925265193, + 0.00463799899443984, + 0.0031645195558667183, + 0.003502072300761938, + 0.004469136707484722, + 0.003364768112078309, + 0.0026122278068214655, + 0.0032403762452304363, + 0.0035367291420698166, + 0.0035232335794717073, + 0.003269039560109377, + 0.004851423669606447, + 0.005576834548264742, + 0.0032659831922501326, + 0.0030094231478869915, + 0.002666330896317959, + 0.005623308010399342, + 0.006955522578209639, + 0.010552817955613136, + 0.011507927440106869, + 0.01282269787043333, + 0.009406846947968006, + 0.0065091983415186405, + 0.00500002596527338, + 0.012949638068675995, + 0.0080055370926857, + 0.007904993370175362, + 0.007418631110340357, + 0.006736880633980036, + 0.00846910197287798, + 0.007016950286924839, + 0.005211787298321724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.023298904299736023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988786101341248 + ], + "opposite_logprob_tokens": [ + -1.8048878908157349 + ], + "logprob_reward": -0.17988786101341248, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988786101341248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10075420141220093, + 0.04896262288093567, + 0.30783170461654663, + 0.0837334468960762, + 0.12924064695835114, + 0.11742730438709259, + 0.03355342149734497, + 0.07347586005926132, + 0.0516459196805954, + 0.05337489768862724 + ], + "credit_map": [ + 0.017483625560998917, + 0.014331362210214138, + 0.020663805305957794, + 0.011825323104858398, + 0.00873622577637434, + 0.007824256084859371, + 0.007044680882245302, + 0.012844925746321678, + 0.0060599008575081825, + 0.007329864427447319, + 0.009769326075911522, + 0.005387131590396166, + 0.005954114254564047, + 0.003951027058064938, + 0.005334251560270786, + 0.005177006591111422, + 0.059022605419158936, + 0.03375694528222084, + 0.033826056867837906, + 0.04430365189909935, + 0.026534441858530045, + 0.03581782057881355, + 0.03662418574094772, + 0.03794598951935768, + 0.015725335106253624, + 0.013956134207546711, + 0.0069100079126656055, + 0.006921416614204645, + 0.007315425667911768, + 0.013102935627102852, + 0.007113019935786724, + 0.012689167633652687, + 0.014038573019206524, + 0.013583295978605747, + 0.021050097420811653, + 0.017979612573981285, + 0.010907449759542942, + 0.01595933549106121, + 0.015497041866183281, + 0.020225247368216515, + 0.011051995679736137, + 0.020028235390782356, + 0.02315434440970421, + 0.01426326297223568, + 0.01139390841126442, + 0.008976944722235203, + 0.011839663609862328, + 0.016718946397304535, + 0.0036468857433646917, + 0.005806020461022854, + 0.0035574068315327168, + 0.0040048412047326565, + 0.0055816262029111385, + 0.003464613575488329, + 0.0031311793718487024, + 0.004360847175121307, + 0.007134927436709404, + 0.006426352076232433, + 0.007703328039497137, + 0.009883470833301544, + 0.0068034701980650425, + 0.019646892324090004, + 0.00755938608199358, + 0.008318030275404453, + 0.004742885008454323, + 0.005965635646134615, + 0.005252813920378685, + 0.005626454949378967, + 0.009866305626928806, + 0.008406316861510277, + 0.006613068748265505, + 0.005172437522560358, + 0.004419747740030289, + 0.005545005202293396, + 0.0064644478261470795, + 0.006552787963300943, + 0.009656645357608795, + 0.006179887801408768, + 0.006646074820309877, + 0.007910303771495819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.17988786101341248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1089036762714386, + 0.03292010724544525, + 0.15781381726264954, + 0.15892770886421204, + 0.10661592334508896, + 0.07359251379966736, + 0.02380278892815113, + 0.09909400343894958, + 0.11967451125383377, + 0.11865495890378952 + ], + "credit_map": [ + 0.012466052547097206, + 0.01554519310593605, + 0.015065350569784641, + 0.014530245214700699, + 0.011347068473696709, + 0.011647414416074753, + 0.013361572287976742, + 0.014940776862204075, + 0.0029034693725407124, + 0.00359169440343976, + 0.0060310158878564835, + 0.004738024901598692, + 0.003586125560104847, + 0.0023400629870593548, + 0.004861664492636919, + 0.00486805010586977, + 0.03376860171556473, + 0.016797978430986404, + 0.017264138907194138, + 0.020454145967960358, + 0.012197742238640785, + 0.019969236105680466, + 0.017254184931516647, + 0.020107798278331757, + 0.03764097020030022, + 0.02038852870464325, + 0.018758883699774742, + 0.014087677001953125, + 0.013757998123764992, + 0.022770443931221962, + 0.012362094596028328, + 0.019161121919751167, + 0.01395221333950758, + 0.010182125493884087, + 0.021252969279885292, + 0.015153978019952774, + 0.01045397762209177, + 0.0055904872715473175, + 0.015032777562737465, + 0.014997385442256927, + 0.0066705066710710526, + 0.011367042548954487, + 0.013978013768792152, + 0.00744371535256505, + 0.01151426974684, + 0.0059579540975391865, + 0.007497464306652546, + 0.009163550101220608, + 0.0020552719943225384, + 0.0033760094083845615, + 0.002661143895238638, + 0.003950465004891157, + 0.0028669964522123337, + 0.0028451920952647924, + 0.002619146602228284, + 0.0034285623114556074, + 0.010570848360657692, + 0.0064405882731080055, + 0.010163376107811928, + 0.012903585098683834, + 0.008291204459965229, + 0.027945194393396378, + 0.010763317346572876, + 0.012015891261398792, + 0.01024357695132494, + 0.015882223844528198, + 0.013619364239275455, + 0.01052620355039835, + 0.034881554543972015, + 0.011961568146944046, + 0.012452064082026482, + 0.010107954032719135, + 0.01043123658746481, + 0.009937001392245293, + 0.007876056246459484, + 0.0206440556794405, + 0.034961163997650146, + 0.009113945998251438, + 0.012851866893470287, + 0.012839624658226967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.011076156049966812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from a slightly upward angle, capturing an elderly woman with silver hair tied in a simple bun, nestled in a tranquil Chinese courtyard in late autumn. She wears a dark blue, double-breasted, coarse cloth training suit, the fabric appearing worn yet crisp and neat. She exudes a timeless calmness. Golden ginkgo leaves and crimson maple leaves slowly fall from sparse branches, forming a thick, vibrant carpet beneath her feet. Her hands steady a pristine longsword, its hilt clad in dark wood, aged by time. The brass guard gleams with a dull luster in the afternoon sun. She lowers herself, her center of gravity steady, her gaze fixed firmly ahead. With a slow, restrained, and powerful turn, the ancient sword traces a steady, flowing arc through the air, its wind gently sweeping up a few fallen leaves, swirling with them. At this point, the camera begins to move in very slowly and smoothly, gradually blurring the background courtyard and focusing entirely on her. Finally, the camera lingers on a close-up of her weathered face, each deep wrinkle telling a long story. Her slightly cloudy yet eagle-like eyes gleam with unyielding focus. The entire scene, bathed in soft, warm light, creates a tranquil atmosphere of solemnity, focus, and power.", + "seed": 14141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14141_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.059707017987966536, + "vlm_logprob": -0.059707017987966536, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.059707017987966536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051105204969644547 + ], + "opposite_logprob_tokens": [ + -3.051105260848999 + ], + "logprob_reward": -0.051105204969644547, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051105204969644547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14107728004455566, + 0.028074150905013084, + 0.04999081790447235, + 0.032139040529727936, + 0.027178440243005753, + 0.124855637550354, + 0.14178873598575592, + 0.07858369499444962, + 0.18988758325576782, + 0.18642455339431763 + ], + "credit_map": [ + 0.015578889288008213, + 0.018015814945101738, + 0.043405141681432724, + 0.009930788539350033, + 0.008903780952095985, + 0.00555839017033577, + 0.01147471647709608, + 0.02820977009832859, + 0.002814078703522682, + 0.00499191926792264, + 0.003335190936923027, + 0.0034821471199393272, + 0.003976393491029739, + 0.0024100071750581264, + 0.003978886641561985, + 0.0030855280347168446, + 0.004702716134488583, + 0.004561915528029203, + 0.006787709891796112, + 0.006407198496162891, + 0.010029725730419159, + 0.005679853726178408, + 0.0071502928622066975, + 0.004671407863497734, + 0.003781126346439123, + 0.0031523960642516613, + 0.008668895810842514, + 0.0023666704073548317, + 0.005062715616077185, + 0.0022590633016079664, + 0.004142665769904852, + 0.002705508377403021, + 0.0033527114428579807, + 0.0030140099115669727, + 0.005556263029575348, + 0.0037679916713386774, + 0.002898715203627944, + 0.0020030306186527014, + 0.004100595135241747, + 0.0024851257912814617, + 0.00952761061489582, + 0.010123010724782944, + 0.010105939581990242, + 0.011106738820672035, + 0.032026197761297226, + 0.01115073449909687, + 0.022378383204340935, + 0.01843702606856823, + 0.00981275737285614, + 0.015057828277349472, + 0.020376458764076233, + 0.03481096774339676, + 0.021638911217451096, + 0.009691381826996803, + 0.020299972966313362, + 0.01010047271847725, + 0.0059792036190629005, + 0.008309852331876755, + 0.006081085652112961, + 0.00968458317220211, + 0.018990829586982727, + 0.007258149795234203, + 0.012261494062840939, + 0.010018502362072468, + 0.017930107191205025, + 0.018526997417211533, + 0.011656415648758411, + 0.02214486338198185, + 0.046994928270578384, + 0.01800210028886795, + 0.035097386687994, + 0.019534800201654434, + 0.016469009220600128, + 0.01654575951397419, + 0.017266951501369476, + 0.022080717608332634, + 0.04665488377213478, + 0.015254550613462925, + 0.037130653858184814, + 0.015022044070065022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.051105204969644547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366978883743286 + ], + "opposite_logprob_tokens": [ + -3.651366949081421 + ], + "logprob_reward": -0.026366978883743286, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366978883743286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19200868904590607, + 0.08919059485197067, + 0.06438183039426804, + 0.05789938569068909, + 0.04374174773693085, + 0.03693968802690506, + 0.1491466611623764, + 0.11559230834245682, + 0.15707437694072723, + 0.09402469545602798 + ], + "credit_map": [ + 0.020596737042069435, + 0.01957380212843418, + 0.0610288642346859, + 0.014990294352173805, + 0.012572011910378933, + 0.012722228653728962, + 0.020755821838974953, + 0.0297689251601696, + 0.009417694993317127, + 0.004436555318534374, + 0.032798320055007935, + 0.009253331460058689, + 0.006455578841269016, + 0.00396429980173707, + 0.006521962583065033, + 0.016342854127287865, + 0.006787039339542389, + 0.0048572001978755, + 0.011252831667661667, + 0.00983564555644989, + 0.008557631634175777, + 0.006565167102962732, + 0.008907962590456009, + 0.007618356496095657, + 0.006441531702876091, + 0.005155200604349375, + 0.007247001398354769, + 0.007073833141475916, + 0.013504857197403908, + 0.005542993079870939, + 0.007249156478792429, + 0.005684814415872097, + 0.0038448451086878777, + 0.004385289736092091, + 0.00736590288579464, + 0.005529170390218496, + 0.007039930671453476, + 0.0036196094006299973, + 0.0077819800935685635, + 0.0041750166565179825, + 0.0044088237918913364, + 0.003569145454093814, + 0.005157713778316975, + 0.0032845563255250454, + 0.006990736350417137, + 0.0029690060764551163, + 0.0056766001507639885, + 0.004883106332272291, + 0.012068085372447968, + 0.013712242245674133, + 0.025196466594934464, + 0.0381883829832077, + 0.012905110605061054, + 0.011996089480817318, + 0.0228438600897789, + 0.012236427515745163, + 0.012834628112614155, + 0.010719596408307552, + 0.0243495162576437, + 0.01761612296104431, + 0.01377648115158081, + 0.01014174334704876, + 0.014196214266121387, + 0.01195800956338644, + 0.02699994295835495, + 0.01710493676364422, + 0.01705090142786503, + 0.03344707563519478, + 0.019973179325461388, + 0.014651256613433361, + 0.01618357188999653, + 0.011663531884551048, + 0.010369956493377686, + 0.010998089797794819, + 0.012570435181260109, + 0.01752896048128605, + 0.011364028789103031, + 0.006947839166969061, + 0.014542958699166775, + 0.009702426381409168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.026366978883743286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594218745827675 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594218745827675, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594218745827675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2513328194618225, + 0.07240257412195206, + 0.04631585627794266, + 0.04699559137225151, + 0.07595851272344589, + 0.04362894594669342, + 0.15902112424373627, + 0.07645292580127716, + 0.11128903925418854, + 0.11660262197256088 + ], + "credit_map": [ + 0.02700112946331501, + 0.015336429700255394, + 0.08606147021055222, + 0.01986796222627163, + 0.017729556187987328, + 0.014579755254089832, + 0.020812010392546654, + 0.04994450509548187, + 0.006870300509035587, + 0.003623227821663022, + 0.027241509407758713, + 0.007191717159003019, + 0.00398681266233325, + 0.0032756708096712828, + 0.003772664349526167, + 0.016440672799944878, + 0.008772228844463825, + 0.004298481624573469, + 0.0058526210486888885, + 0.004511876031756401, + 0.006758313626050949, + 0.004572094418108463, + 0.005910841282457113, + 0.005639403127133846, + 0.005770890507847071, + 0.005475254263728857, + 0.006045167800039053, + 0.005753541365265846, + 0.00754329701885581, + 0.0048929862678050995, + 0.005160937551409006, + 0.006353522185236216, + 0.011538531631231308, + 0.008160009980201721, + 0.016273045912384987, + 0.007294138427823782, + 0.008979495614767075, + 0.00487874960526824, + 0.010604447685182095, + 0.00823009479790926, + 0.0057578361593186855, + 0.0038912983145564795, + 0.006097510922700167, + 0.00421776482835412, + 0.004679771605879068, + 0.004366611130535603, + 0.006358468439429998, + 0.00825969036668539, + 0.012229478918015957, + 0.013750995509326458, + 0.030984239652752876, + 0.03780113533139229, + 0.012639597058296204, + 0.012238983064889908, + 0.024863354861736298, + 0.014513352885842323, + 0.008553411811590195, + 0.007281117606908083, + 0.015315111726522446, + 0.01224443968385458, + 0.009798643179237843, + 0.007530493661761284, + 0.00873072911053896, + 0.006998986471444368, + 0.013239413499832153, + 0.011279620230197906, + 0.01711432822048664, + 0.021261177957057953, + 0.013522022403776646, + 0.01099502481520176, + 0.015086849220097065, + 0.008790605701506138, + 0.012557530775666237, + 0.0119025819003582, + 0.018186043947935104, + 0.02390890382230282, + 0.014557991176843643, + 0.00885776150971651, + 0.01566830277442932, + 0.010963508859276772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.020594218745827675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988964915275574 + ], + "opposite_logprob_tokens": [ + -1.8048896789550781 + ], + "logprob_reward": -0.17988964915275574, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988964915275574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576838195323944, + 0.074236661195755, + 0.04433926194906235, + 0.06842238456010818, + 0.07872430235147476, + 0.026308804750442505, + 0.2434958815574646, + 0.11191204935312271, + 0.07517365366220474, + 0.11970314383506775 + ], + "credit_map": [ + 0.018275845795869827, + 0.02065262570977211, + 0.03977185860276222, + 0.01591270975768566, + 0.011866888962686062, + 0.012214825488626957, + 0.015136620961129665, + 0.023852458223700523, + 0.0070609175600111485, + 0.0031733978539705276, + 0.028492972254753113, + 0.007898973301053047, + 0.004972249269485474, + 0.003097587265074253, + 0.004329910036176443, + 0.015210656449198723, + 0.0061237383633852005, + 0.005789789836853743, + 0.006401999853551388, + 0.005958572030067444, + 0.005748431198298931, + 0.004730090033262968, + 0.005550650879740715, + 0.00403599301353097, + 0.009003369137644768, + 0.00548788346350193, + 0.01572447083890438, + 0.006516861263662577, + 0.010914428159594536, + 0.005208232905715704, + 0.007522568106651306, + 0.008044575341045856, + 0.008276092819869518, + 0.00710633397102356, + 0.013894418254494667, + 0.010417016223073006, + 0.008946557529270649, + 0.00574722932651639, + 0.011539401486515999, + 0.012797252275049686, + 0.002869199262931943, + 0.002009544987231493, + 0.003957964479923248, + 0.002581789158284664, + 0.0030884139705449343, + 0.0015250667929649353, + 0.004349665716290474, + 0.005927163176238537, + 0.02172323875129223, + 0.02526487223803997, + 0.042174331843853, + 0.06058136746287346, + 0.019638827070593834, + 0.020503917708992958, + 0.03636642545461655, + 0.017242930829524994, + 0.014055890031158924, + 0.010839467868208885, + 0.010169057175517082, + 0.018096452578902245, + 0.017302367836236954, + 0.009242173284292221, + 0.018737876787781715, + 0.013468771241605282, + 0.008698425255715847, + 0.007093682885169983, + 0.010686365887522697, + 0.012898053042590618, + 0.01282693911343813, + 0.0062172794714570045, + 0.009055595844984055, + 0.007697318680584431, + 0.01201535202562809, + 0.011288302950561047, + 0.019278543069958687, + 0.013961824588477612, + 0.018857751041650772, + 0.013576338067650795, + 0.01811067759990692, + 0.012614358216524124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.17988964915275574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579038187861443 + ], + "opposite_logprob_tokens": [ + -3.8955790996551514 + ], + "logprob_reward": -0.020579038187861443, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579038187861443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19721953570842743, + 0.05750924348831177, + 0.06471581757068634, + 0.046976882964372635, + 0.04445818439126015, + 0.05940597504377365, + 0.08898486196994781, + 0.13230019807815552, + 0.08607379347085953, + 0.22235549986362457 + ], + "credit_map": [ + 0.021158359944820404, + 0.02094373293220997, + 0.061549410223960876, + 0.013948401436209679, + 0.014333151280879974, + 0.012271462939679623, + 0.01828853413462639, + 0.03472648933529854, + 0.005028238985687494, + 0.004088342189788818, + 0.016421880573034286, + 0.008259128779172897, + 0.005019183270633221, + 0.0035060143563896418, + 0.004212372470647097, + 0.010974080301821232, + 0.005831113085150719, + 0.005578096490353346, + 0.008716513402760029, + 0.0066421800293028355, + 0.01539697963744402, + 0.005425552371889353, + 0.011466843076050282, + 0.005658534821122885, + 0.005253495182842016, + 0.004328136797994375, + 0.010937556624412537, + 0.005476239137351513, + 0.005694519262760878, + 0.00420547416433692, + 0.005738670006394386, + 0.005342795047909021, + 0.00593227706849575, + 0.0039854575879871845, + 0.0076387072913348675, + 0.003917901311069727, + 0.008005794137716293, + 0.00361051713116467, + 0.007478301413357258, + 0.0038892331067472696, + 0.005621416959911585, + 0.004810982849448919, + 0.008041338063776493, + 0.005727745592594147, + 0.013936922885477543, + 0.007778212893754244, + 0.007230882532894611, + 0.006258473265916109, + 0.006422997452318668, + 0.00792725756764412, + 0.01335059106349945, + 0.020793674513697624, + 0.012643183581531048, + 0.007030673790723085, + 0.013059199787676334, + 0.007757293991744518, + 0.012459744699299335, + 0.01041323971003294, + 0.019597237929701805, + 0.0211747158318758, + 0.017260489985346794, + 0.009291631169617176, + 0.021360062062740326, + 0.020743070170283318, + 0.011216371320188046, + 0.008524605073034763, + 0.010557367466390133, + 0.013941965065896511, + 0.013984257355332375, + 0.007583459839224815, + 0.012140109203755856, + 0.008125664666295052, + 0.026197688654065132, + 0.02281365916132927, + 0.04277600720524788, + 0.036299169063568115, + 0.03365449234843254, + 0.016614504158496857, + 0.026599233970046043, + 0.017400747165083885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.020579038187861443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, shot from a slightly upward angle, capturing an elderly woman with silver hair tied in a simple bun, nestled in a tranquil Chinese courtyard in late autumn. She wears a dark blue, double-breasted, coarse cloth training suit, the fabric appearing worn yet crisp and neat. She exudes a timeless calmness. Golden ginkgo leaves and crimson maple leaves slowly fall from sparse branches, forming a thick, vibrant carpet beneath her feet. Her hands steady a pristine longsword, its hilt clad in dark wood, aged by time. The brass guard gleams with a dull luster in the afternoon sun. She lowers herself, her center of gravity steady, her gaze fixed firmly ahead. With a slow, restrained, and powerful turn, the ancient sword traces a steady, flowing arc through the air, its wind gently sweeping up a few fallen leaves, swirling with them. At this point, the camera begins to move in very slowly and smoothly, gradually blurring the background courtyard and focusing entirely on her. Finally, the camera lingers on a close-up of her weathered face, each deep wrinkle telling a long story. Her slightly cloudy yet eagle-like eyes gleam with unyielding focus. The entire scene, bathed in soft, warm light, creates a tranquil atmosphere of solemnity, focus, and power.", + "seed": 14142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_2/index_1691_seed_14142_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12134179342538118, + "vlm_logprob": -0.12134179342538118, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.12134179342538118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07218042016029358 + ], + "opposite_logprob_tokens": [ + -2.6971805095672607 + ], + "logprob_reward": -0.07218042016029358, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07218042016029358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04652485251426697, + 0.02452397346496582, + 0.05849352106451988, + 0.02265426144003868, + 0.16370633244514465, + 0.5117990970611572, + 0.014388227835297585, + 0.02163943648338318, + 0.09806758910417557, + 0.038202714174985886 + ], + "credit_map": [ + 0.007764142472296953, + 0.004221722017973661, + 0.005714491941034794, + 0.005878563039004803, + 0.006529562175273895, + 0.00526457978412509, + 0.005706456024199724, + 0.005445334129035473, + 0.003123116446658969, + 0.0033885117154568434, + 0.0023141687270253897, + 0.003352782689034939, + 0.004160510376095772, + 0.0027545010671019554, + 0.003275517839938402, + 0.002154863439500332, + 0.002929633716121316, + 0.002977138850837946, + 0.01320418156683445, + 0.023858459666371346, + 0.0019192983163520694, + 0.0036763995885849, + 0.006018776446580887, + 0.003909627441316843, + 0.0030859995167702436, + 0.003911519888788462, + 0.004928299691528082, + 0.0017302468186244369, + 0.002384857041761279, + 0.0023140041157603264, + 0.0022707991302013397, + 0.0020285374484956264, + 0.014891949482262135, + 0.04363897815346718, + 0.020356815308332443, + 0.017448846250772476, + 0.01617935113608837, + 0.02252309024333954, + 0.014552880078554153, + 0.014114419929683208, + 0.04610295966267586, + 0.14715927839279175, + 0.11300297826528549, + 0.038850996643304825, + 0.05381312966346741, + 0.0369625985622406, + 0.03728312999010086, + 0.03862404450774193, + 0.0019199171802029014, + 0.0013025844236835837, + 0.0017274681013077497, + 0.002111748093739152, + 0.002065537963062525, + 0.0019360259175300598, + 0.0015533962287008762, + 0.0017715508583933115, + 0.002290894743055105, + 0.0033475507516413927, + 0.002476150868460536, + 0.0015011350624263287, + 0.003866961458697915, + 0.0042676241137087345, + 0.0017188129713758826, + 0.0021703066304326057, + 0.006488624960184097, + 0.01891512982547283, + 0.009579465724527836, + 0.009109134785830975, + 0.01997840404510498, + 0.017270684242248535, + 0.008598550222814083, + 0.008127597160637379, + 0.0027544174809008837, + 0.008699040859937668, + 0.0029992542695254087, + 0.005238272249698639, + 0.0070302230305969715, + 0.0034946114756166935, + 0.004889755044132471, + 0.0030971390660852194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.07218042016029358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829099774360657 + ], + "opposite_logprob_tokens": [ + -3.5298290252685547 + ], + "logprob_reward": -0.029829099774360657, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829099774360657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0867667943239212, + 0.022671349346637726, + 0.05693656951189041, + 0.02242913283407688, + 0.08491206169128418, + 0.5520563125610352, + 0.018669450655579567, + 0.07879704982042313, + 0.04014645516872406, + 0.03661482408642769 + ], + "credit_map": [ + 0.015131115913391113, + 0.006051211152225733, + 0.010770082473754883, + 0.012989537790417671, + 0.01139266137033701, + 0.011504076421260834, + 0.008207090198993683, + 0.010721016675233841, + 0.0021499888971447945, + 0.0020695242565125227, + 0.0026289618108421564, + 0.003288750071078539, + 0.004261746536940336, + 0.003064914606511593, + 0.003074159612879157, + 0.002133304253220558, + 0.0028022583574056625, + 0.0037840104196220636, + 0.017701486125588417, + 0.017583239823579788, + 0.002592303091660142, + 0.0033307259436696768, + 0.004778859671205282, + 0.004363685380667448, + 0.0017055042553693056, + 0.0031317458488047123, + 0.0045730918645858765, + 0.002124173566699028, + 0.00445148441940546, + 0.0024008157197386026, + 0.0023860251531004906, + 0.001656293054111302, + 0.007228207774460316, + 0.0213725958019495, + 0.011434274725615978, + 0.008764606900513172, + 0.009087211452424526, + 0.011187318712472916, + 0.007764854002743959, + 0.008072994649410248, + 0.048933010548353195, + 0.1567971259355545, + 0.10957041382789612, + 0.04702053964138031, + 0.058118630200624466, + 0.042692575603723526, + 0.04317291080951691, + 0.045751068741083145, + 0.0012513617984950542, + 0.002172339940443635, + 0.001635029911994934, + 0.0023762064520269632, + 0.002707927254959941, + 0.0032334851566702127, + 0.003654778702184558, + 0.001638322719372809, + 0.006771364714950323, + 0.009573790244758129, + 0.01108898688107729, + 0.010354923084378242, + 0.011424127034842968, + 0.01178867369890213, + 0.008999130688607693, + 0.008796051144599915, + 0.002505310345441103, + 0.0073857796378433704, + 0.004017659928649664, + 0.0036099539138376713, + 0.010208426974713802, + 0.00492183119058609, + 0.003570803441107273, + 0.003926691133528948, + 0.0022687665186822414, + 0.0074700466357171535, + 0.003730203490704298, + 0.004710161592811346, + 0.004863522481173277, + 0.004072458017617464, + 0.0059821708127856255, + 0.0035174950025975704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.029829099774360657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15455588698387146, + 0.04025910794734955, + 0.21949775516986847, + 0.02225341647863388, + 0.14175181090831757, + 0.18479223549365997, + 0.03530680760741234, + 0.047521963715553284, + 0.1019720658659935, + 0.05208887159824371 + ], + "credit_map": [ + 0.02100357785820961, + 0.015055433847010136, + 0.01837979257106781, + 0.019508440047502518, + 0.022622590884566307, + 0.027721794322133064, + 0.01104077510535717, + 0.01922348327934742, + 0.0038051248993724585, + 0.005694892257452011, + 0.0075307125225663185, + 0.004305660724639893, + 0.004737086594104767, + 0.006347682327032089, + 0.0030873939394950867, + 0.004750553052872419, + 0.01611556112766266, + 0.026269620284438133, + 0.023161374032497406, + 0.05667634308338165, + 0.012894924730062485, + 0.019622813910245895, + 0.016028858721256256, + 0.04872828349471092, + 0.002604134613648057, + 0.0022996612824499607, + 0.003526979358866811, + 0.0018144816858693957, + 0.003073721891269088, + 0.0033552588429301977, + 0.002650141017511487, + 0.0029290386009961367, + 0.012854366563260555, + 0.03632622957229614, + 0.01627393253147602, + 0.015015799552202225, + 0.01485169492661953, + 0.020496264100074768, + 0.012450831942260265, + 0.01348270382732153, + 0.01612189970910549, + 0.04414341598749161, + 0.04601795971393585, + 0.015071488916873932, + 0.015096388757228851, + 0.014392031356692314, + 0.01670130528509617, + 0.017247751355171204, + 0.0026835233438760042, + 0.0028309873305261135, + 0.0038130474276840687, + 0.005643640644848347, + 0.0040920330211520195, + 0.006800482515245676, + 0.006382637191563845, + 0.003060455434024334, + 0.0034970175474882126, + 0.005357625428587198, + 0.004900806583464146, + 0.003831275971606374, + 0.01015458069741726, + 0.008332117460668087, + 0.006210905034095049, + 0.005237634759396315, + 0.006642581894993782, + 0.014106519520282745, + 0.00980412494391203, + 0.007258186582475901, + 0.025918832048773766, + 0.021754253655672073, + 0.006580321118235588, + 0.009907247498631477, + 0.004930319730192423, + 0.006933695636689663, + 0.0064171599224209785, + 0.0049355486407876015, + 0.00766014913097024, + 0.00664930883795023, + 0.007122750394046307, + 0.007439938373863697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743255376815796 + ], + "opposite_logprob_tokens": [ + -0.9743255376815796 + ], + "logprob_reward": -0.4743255376815796, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743255376815796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14092056453227997, + 0.05270042270421982, + 0.07207435369491577, + 0.016231494024395943, + 0.17970848083496094, + 0.3711162507534027, + 0.030834630131721497, + 0.05021575838327408, + 0.039541199803352356, + 0.04665687307715416 + ], + "credit_map": [ + 0.0130165359005332, + 0.014014458283782005, + 0.016750916838645935, + 0.021353721618652344, + 0.024104245007038116, + 0.018221385776996613, + 0.017118418589234352, + 0.01634088344871998, + 0.005230297334492207, + 0.005122709088027477, + 0.005988197401165962, + 0.008188114501535892, + 0.009826639667153358, + 0.006338808219879866, + 0.007551094982773066, + 0.004454560112208128, + 0.004565803799778223, + 0.006031425204128027, + 0.0068419622257351875, + 0.02595684677362442, + 0.004086669068783522, + 0.006189348641782999, + 0.0068288943730294704, + 0.011573399417102337, + 0.0015017581172287464, + 0.002586043905466795, + 0.0027771920431405306, + 0.0016112662851810455, + 0.0028884829953312874, + 0.0018761511892080307, + 0.001720851520076394, + 0.0012697477359324694, + 0.016275571659207344, + 0.04659412428736687, + 0.021817229688167572, + 0.01838671788573265, + 0.018814515322446823, + 0.025678860023617744, + 0.015876861289143562, + 0.016264598816633224, + 0.02998870238661766, + 0.09750759601593018, + 0.08669926971197128, + 0.03124810755252838, + 0.033749114722013474, + 0.028697853907942772, + 0.029444975778460503, + 0.03378064185380936, + 0.00186284095980227, + 0.006410786882042885, + 0.002878935541957617, + 0.004390712361782789, + 0.006317494437098503, + 0.0032249956857413054, + 0.0034422622993588448, + 0.0023065998684614897, + 0.003594862762838602, + 0.0057811737060546875, + 0.004402055870741606, + 0.00430363928899169, + 0.011077912524342537, + 0.009944260120391846, + 0.005529970396310091, + 0.0055818865075707436, + 0.0026695462875068188, + 0.005484841763973236, + 0.003955237101763487, + 0.0031146726105362177, + 0.010479393415153027, + 0.007244251202791929, + 0.002685217186808586, + 0.003908040467649698, + 0.0032900203950703144, + 0.0071516805328428745, + 0.005544471554458141, + 0.0042261527851223946, + 0.005706390365958214, + 0.008927576243877411, + 0.005977632012218237, + 0.005832951981574297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.4743255376815796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly woman with silver hair tied in a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue, double-breasted, coarse cloth training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a Chinese courtyard with falling ginkgo and maple leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman perform a slow turn with a longsword that sweeps up fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot from an upward angle and end with a close-up of her face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784595109522343 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784595109522343, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784595109522343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0353594608604908, + 0.01326137688010931, + 0.050332214683294296, + 0.026123806834220886, + 0.2833327651023865, + 0.5067419409751892, + 0.014808455482125282, + 0.03408915549516678, + 0.019796444103121758, + 0.016154401004314423 + ], + "credit_map": [ + 0.005749073810875416, + 0.0053342180326581, + 0.0040169223211705685, + 0.0030907918699085712, + 0.005241020582616329, + 0.0048573752865195274, + 0.003334733424708247, + 0.0037353262305259705, + 0.0010031990241259336, + 0.0015651843277737498, + 0.0015055523253977299, + 0.002244223142042756, + 0.001640290836803615, + 0.0018535911804065108, + 0.0016969118732959032, + 0.0017524241702631116, + 0.0031959896441549063, + 0.004241610411554575, + 0.004601558670401573, + 0.015384769067168236, + 0.002912994707003236, + 0.003950934857130051, + 0.004149207379668951, + 0.011895151808857918, + 0.002355545060709119, + 0.005487917456775904, + 0.004655991215258837, + 0.0021335617639124393, + 0.002759503433480859, + 0.0038831327110528946, + 0.002286739181727171, + 0.0025614160113036633, + 0.0247056744992733, + 0.07250577956438065, + 0.034048184752464294, + 0.02925645187497139, + 0.03179392218589783, + 0.03929625079035759, + 0.025155754759907722, + 0.0265707578510046, + 0.041664063930511475, + 0.13835550844669342, + 0.11800597608089447, + 0.04217655584216118, + 0.04769930616021156, + 0.037534769624471664, + 0.03886207193136215, + 0.0424436554312706, + 0.0015037702396512032, + 0.0015426974277943373, + 0.0014096583472564816, + 0.0021050814539194107, + 0.0022857035510241985, + 0.0019666326697915792, + 0.002467677928507328, + 0.001527235028333962, + 0.002717462368309498, + 0.004058703780174255, + 0.003642579074949026, + 0.0029175698291510344, + 0.006355182733386755, + 0.006778760347515345, + 0.0038184532895684242, + 0.003800446167588234, + 0.0014236059505492449, + 0.0031977961771190166, + 0.002162980381399393, + 0.001907415222376585, + 0.003896882990375161, + 0.0031184658873826265, + 0.0016570178559049964, + 0.002432281617075205, + 0.0009821662679314613, + 0.0025683969724923372, + 0.002070972230285406, + 0.001873921137303114, + 0.0024700057692825794, + 0.0022022901102900505, + 0.002366140717640519, + 0.0016205089632421732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.009784595109522343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_tensor([4208, 4208])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_tensor([4208, 4208])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ceb6cd3025cbf001b647002e9680d48fc8a24223 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_tensor([4208, 4208])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, long, overhead shot, slowly tilting upward from street level, plunges the viewer into a cyberpunk future drenched in perpetual acid rain and shimmering holographic advertisements. Fine acid rain falls like a curtain, creating tiny ripples on the slippery asphalt, while massive, volatile neon advertisements projected between the glass facades of skyscrapers cast eerie, hypersaturated blues, purples, and pinks against the gloomy night sky. Against this chaotic backdrop, the colossal Godzilla clashes fiercely with a comparable mechanical behemoth. Godzilla's rough, rock-gray scales gleam wetly in the rain, an ominous atomic blue shimmering through the gaps between them. Godzilla leans forward slightly, muscles tense, brimming with primal fury. The opposing behemoth is constructed of polished titanium and carbon fiber, its geometric armor gleaming with a cold metallic sheen. Complex hydraulic lines and glowing red sensors crisscross its body, its shoulder-mounted weapon systems activated. Suddenly, Godzilla's dorsal fins began to light up one by one, emitting a pulsating blue light, and a hot atomic breath suddenly burst out from its open jaws. At the same time, the core reactor in the chest of the mechanical monster burst into a dazzling white light, and a precise beam of plasma energy shot towards it. The two destructive energies collided violently in mid-air, producing a spherical, distorted energy explosion, which instantly split the gloomy sky and illuminated the entire city, creating a tense, grand and doomsday atmosphere of epic confrontation.", + "seed": 15456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15456_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2347784917568788, + "vlm_logprob": -0.2347784917568788, + "vlm_yes_no_margin": 3.925000011920929, + "vlm_token_credit": -0.2347784917568788, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0860381126403809 + ], + "opposite_logprob_tokens": [ + -0.46103817224502563 + ], + "logprob_reward": -1.0860381126403809, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0860381126403809, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.27787232398986816, + 0.27285024523735046, + 0.07155606150627136, + 0.03790158033370972, + 0.06529044359922409, + 0.044120267033576965, + 0.058403078466653824, + 0.0327262356877327, + 0.08087944239377975, + 0.05840035155415535 + ], + "credit_map": [ + 0.017061954364180565, + 0.021588722243905067, + 0.02649036981165409, + 0.05402771383523941, + 0.03365911915898323, + 0.02783939614892006, + 0.06909624487161636, + 0.028108801692724228, + 0.02211136184632778, + 0.03840884938836098, + 0.03448615223169327, + 0.031437307596206665, + 0.046765636652708054, + 0.033951833844184875, + 0.028269005939364433, + 0.03742009028792381, + 0.010883604176342487, + 0.01071907952427864, + 0.006958639249205589, + 0.008652601391077042, + 0.010651765391230583, + 0.009261556901037693, + 0.007899675518274307, + 0.00652913935482502, + 0.00377817009575665, + 0.0030798502266407013, + 0.004037933424115181, + 0.005121281836181879, + 0.0046461704187095165, + 0.003996890503913164, + 0.008039012551307678, + 0.005202271975576878, + 0.00756172277033329, + 0.0063766976818442345, + 0.010866803117096424, + 0.008408532477915287, + 0.008174885995686054, + 0.008015305735170841, + 0.007207175251096487, + 0.008679324761033058, + 0.00464383140206337, + 0.005229069851338863, + 0.006408125162124634, + 0.004752822685986757, + 0.0045803808607161045, + 0.004378718324005604, + 0.006505636032670736, + 0.0076216827146708965, + 0.004514207132160664, + 0.008924426510930061, + 0.007822261191904545, + 0.006449394393712282, + 0.006274580489844084, + 0.004781061317771673, + 0.007574055343866348, + 0.01206309162080288, + 0.003520206781104207, + 0.004714986775070429, + 0.0032434933818876743, + 0.0070671518333256245, + 0.0033091166988015175, + 0.0022382086608558893, + 0.004030629992485046, + 0.004602442029863596, + 0.0146141042932868, + 0.01418326050043106, + 0.010058077983558178, + 0.0051518711261451244, + 0.005952927749603987, + 0.014084932394325733, + 0.009987007826566696, + 0.0068472567945718765, + 0.005543735809624195, + 0.006974288262426853, + 0.006544357631355524, + 0.006640757899731398, + 0.009633015841245651, + 0.00784087274223566, + 0.008994076400995255, + 0.0062292469665408134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -1.0860381126403809 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904121279716492 + ], + "opposite_logprob_tokens": [ + -2.5790412425994873 + ], + "logprob_reward": -0.07904121279716492, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904121279716492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2308749109506607, + 0.22696584463119507, + 0.07788937538862228, + 0.10098033398389816, + 0.0704057365655899, + 0.06906220316886902, + 0.05981491506099701, + 0.02507266402244568, + 0.09627494215965271, + 0.042659129947423935 + ], + "credit_map": [ + 0.016415925696492195, + 0.015992436558008194, + 0.020806429907679558, + 0.051560405641794205, + 0.02494930848479271, + 0.043676577508449554, + 0.03256706893444061, + 0.024906741455197334, + 0.01763555407524109, + 0.03330538794398308, + 0.028496813029050827, + 0.027264805510640144, + 0.025706855580210686, + 0.03489042818546295, + 0.027671901509165764, + 0.03199407830834389, + 0.00817087385803461, + 0.008468125015497208, + 0.009121928364038467, + 0.007973061874508858, + 0.010700548999011517, + 0.01873907260596752, + 0.007308185566216707, + 0.0074075739830732346, + 0.012627425603568554, + 0.01023196242749691, + 0.012871588580310345, + 0.008853957056999207, + 0.013373956084251404, + 0.012727901339530945, + 0.016312045976519585, + 0.013981481082737446, + 0.007471124175935984, + 0.006067026872187853, + 0.01126341987401247, + 0.005661701317876577, + 0.008770722895860672, + 0.011228499934077263, + 0.008941774256527424, + 0.01100145559757948, + 0.006188588682562113, + 0.009898561984300613, + 0.007169359363615513, + 0.007167183794081211, + 0.012149875052273273, + 0.011308833956718445, + 0.007260468322783709, + 0.007919317111372948, + 0.00813083816319704, + 0.009366434998810291, + 0.00709331501275301, + 0.005218209698796272, + 0.005705848336219788, + 0.004944754298776388, + 0.009562110528349876, + 0.009793400764465332, + 0.0027357772924005985, + 0.0037230353336781263, + 0.0034515606239438057, + 0.003050331026315689, + 0.0027530367951840162, + 0.0033270935527980328, + 0.0028081994969397783, + 0.0032236273400485516, + 0.013617083430290222, + 0.01818242482841015, + 0.011652582325041294, + 0.008192311972379684, + 0.010848178528249264, + 0.010299545712769032, + 0.013776779174804688, + 0.009706028737127781, + 0.00586698530241847, + 0.007314711343497038, + 0.00541450222954154, + 0.0032155150547623634, + 0.003974504768848419, + 0.007130274083465338, + 0.005217813421040773, + 0.004524817690253258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07904121279716492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199935657903552 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003199935657903552, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199935657903552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3895568549633026, + 0.11802902817726135, + 0.047136418521404266, + 0.04732677713036537, + 0.0509818010032177, + 0.0818353220820427, + 0.06090458109974861, + 0.03266028314828873, + 0.07842490822076797, + 0.0931439995765686 + ], + "credit_map": [ + 0.023479288443922997, + 0.02704649791121483, + 0.041046738624572754, + 0.03206206485629082, + 0.056447118520736694, + 0.049083832651376724, + 0.10107263177633286, + 0.059318698942661285, + 0.009175742976367474, + 0.00970778614282608, + 0.014360031113028526, + 0.007761478889733553, + 0.0275141391903162, + 0.018120286986231804, + 0.013668491505086422, + 0.01772107370197773, + 0.005281664431095123, + 0.005183106753975153, + 0.0047449334524571896, + 0.00505903922021389, + 0.007997597567737103, + 0.006435653194785118, + 0.0074457344599068165, + 0.004988689906895161, + 0.004461398348212242, + 0.004012456629425287, + 0.005546466913074255, + 0.004977219272404909, + 0.004675908479839563, + 0.0054167816415429115, + 0.01110332366079092, + 0.0071332198567688465, + 0.005019675008952618, + 0.005298946984112263, + 0.008602033369243145, + 0.005935652181506157, + 0.005086489021778107, + 0.003902266500517726, + 0.00866840872913599, + 0.008468329906463623, + 0.005843260325491428, + 0.012106076814234257, + 0.010099528357386589, + 0.009326671250164509, + 0.007565884850919247, + 0.007697416469454765, + 0.013385402970016003, + 0.01581108383834362, + 0.004901601001620293, + 0.01032397523522377, + 0.00774760264903307, + 0.0077455188147723675, + 0.0075818197801709175, + 0.006695208605378866, + 0.008088434115052223, + 0.007820419035851955, + 0.003302997909486294, + 0.0052044750191271305, + 0.0044658491387963295, + 0.0038786332588642836, + 0.004209890961647034, + 0.003500591963529587, + 0.004813418257981539, + 0.0032844271045178175, + 0.014019675552845001, + 0.014548583887517452, + 0.00951424427330494, + 0.005956796929240227, + 0.00569006847217679, + 0.008944709785282612, + 0.01135281752794981, + 0.008398010395467281, + 0.013472683727741241, + 0.014044790528714657, + 0.01085873693227768, + 0.006845367141067982, + 0.011066996492445469, + 0.013490010052919388, + 0.01260344497859478, + 0.010761967860162258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003199935657903552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015190026024356484 + ], + "opposite_logprob_tokens": [ + -6.501519203186035 + ], + "logprob_reward": -0.0015190026024356484, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015190026024356484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3452891707420349, + 0.1388797163963318, + 0.049167223274707794, + 0.07824045419692993, + 0.07694690674543381, + 0.05766405537724495, + 0.0607406385242939, + 0.02237815037369728, + 0.05853579193353653, + 0.11215785890817642 + ], + "credit_map": [ + 0.012567580677568913, + 0.019973665475845337, + 0.02756931073963642, + 0.023842178285121918, + 0.037634722888469696, + 0.09394910931587219, + 0.08643932640552521, + 0.04331330582499504, + 0.012658583000302315, + 0.02254669740796089, + 0.01732761599123478, + 0.01795855537056923, + 0.020316146314144135, + 0.02014980837702751, + 0.012924442067742348, + 0.0149978743866086, + 0.007798515260219574, + 0.008189612999558449, + 0.004059820901602507, + 0.0058841523714363575, + 0.006881480570882559, + 0.007757469080388546, + 0.004372402094304562, + 0.0042237709276378155, + 0.010850985534489155, + 0.01073560118675232, + 0.006756718270480633, + 0.0077787647023797035, + 0.014203653670847416, + 0.009733976796269417, + 0.008829274214804173, + 0.009351477958261967, + 0.004916986916214228, + 0.004950948059558868, + 0.009573887102305889, + 0.004556082189083099, + 0.01590246520936489, + 0.01809295266866684, + 0.007248551584780216, + 0.011705039069056511, + 0.004858800210058689, + 0.010733116418123245, + 0.006803199648857117, + 0.006621300708502531, + 0.005369147285819054, + 0.0055470820516347885, + 0.009069232270121574, + 0.008662174455821514, + 0.005218717269599438, + 0.010764510370790958, + 0.007326972670853138, + 0.00726957805454731, + 0.008501656353473663, + 0.007107187062501907, + 0.007265530992299318, + 0.007286485750228167, + 0.002149188658222556, + 0.004069229122251272, + 0.002818248700350523, + 0.0034425766207277775, + 0.0026418338529765606, + 0.0021310693118721247, + 0.0029865398537367582, + 0.002139465184882283, + 0.009442899376153946, + 0.009669316932559013, + 0.008057714439928532, + 0.00493545550853014, + 0.007786623202264309, + 0.005665836855769157, + 0.007598082069307566, + 0.005379857961088419, + 0.013076293282210827, + 0.018790654838085175, + 0.011324098333716393, + 0.010145576670765877, + 0.013277005404233932, + 0.012123122811317444, + 0.019825639203190804, + 0.013595471158623695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0015190026024356484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2635251581668854, + 0.09500554203987122, + 0.029795654118061066, + 0.04990347847342491, + 0.18257121741771698, + 0.10665794461965561, + 0.03811589628458023, + 0.047665249556303024, + 0.09105368703603745, + 0.09570623189210892 + ], + "credit_map": [ + 0.012069854885339737, + 0.01558059174567461, + 0.021831965073943138, + 0.023182854056358337, + 0.025956612080335617, + 0.07071678340435028, + 0.05719757080078125, + 0.036988936364650726, + 0.008517567999660969, + 0.012253191322088242, + 0.011359691619873047, + 0.009482492692768574, + 0.018617257475852966, + 0.015071269124746323, + 0.009077729657292366, + 0.010626341216266155, + 0.0039901756681501865, + 0.004596245940774679, + 0.0027883106376975775, + 0.0029712049290537834, + 0.0038687821943312883, + 0.00515894778072834, + 0.003510609734803438, + 0.002911377465352416, + 0.004889577627182007, + 0.004342569503933191, + 0.005062134470790625, + 0.006585337687283754, + 0.006734807509928942, + 0.006256961263716221, + 0.008803587406873703, + 0.007228500675410032, + 0.010394200682640076, + 0.011645706370472908, + 0.02101866714656353, + 0.008554596453905106, + 0.040821634232997894, + 0.04937354475259781, + 0.013154544867575169, + 0.027608312666416168, + 0.006479289848357439, + 0.017134131863713264, + 0.014017047360539436, + 0.008751710876822472, + 0.01576356403529644, + 0.01667501963675022, + 0.012399392202496529, + 0.015437789261341095, + 0.003689625533297658, + 0.004601004999130964, + 0.00454592565074563, + 0.005622156895697117, + 0.004236516077071428, + 0.004219789057970047, + 0.0054805343970656395, + 0.005720340181142092, + 0.0033406587317585945, + 0.010465867817401886, + 0.006106496322900057, + 0.00801074504852295, + 0.00481761246919632, + 0.004703480750322342, + 0.00537160225212574, + 0.004848790820688009, + 0.01864650845527649, + 0.015532981604337692, + 0.010858631692826748, + 0.006025195121765137, + 0.00827487651258707, + 0.013863077387213707, + 0.0095236636698246, + 0.008328759111464024, + 0.01634061709046364, + 0.014169659465551376, + 0.012922101654112339, + 0.005958307068794966, + 0.013737120665609837, + 0.011034327559173107, + 0.012256890535354614, + 0.009287208318710327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.004094195086508989 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long, overhead shot, slowly tilting upward from street level, plunges the viewer into a cyberpunk future drenched in perpetual acid rain and shimmering holographic advertisements. Fine acid rain falls like a curtain, creating tiny ripples on the slippery asphalt, while massive, volatile neon advertisements projected between the glass facades of skyscrapers cast eerie, hypersaturated blues, purples, and pinks against the gloomy night sky. Against this chaotic backdrop, the colossal Godzilla clashes fiercely with a comparable mechanical behemoth. Godzilla's rough, rock-gray scales gleam wetly in the rain, an ominous atomic blue shimmering through the gaps between them. Godzilla leans forward slightly, muscles tense, brimming with primal fury. The opposing behemoth is constructed of polished titanium and carbon fiber, its geometric armor gleaming with a cold metallic sheen. Complex hydraulic lines and glowing red sensors crisscross its body, its shoulder-mounted weapon systems activated. Suddenly, Godzilla's dorsal fins began to light up one by one, emitting a pulsating blue light, and a hot atomic breath suddenly burst out from its open jaws. At the same time, the core reactor in the chest of the mechanical monster burst into a dazzling white light, and a precise beam of plasma energy shot towards it. The two destructive energies collided violently in mid-air, producing a spherical, distorted energy explosion, which instantly split the gloomy sky and illuminated the entire city, creating a tense, grand and doomsday atmosphere of epic confrontation.", + "seed": 15457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15457_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16373199245426803, + "vlm_logprob": -0.16373199245426803, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.16373199245426803, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.778639018535614 + ], + "opposite_logprob_tokens": [ + -0.653639018535614 + ], + "logprob_reward": -0.778639018535614, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.778639018535614, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2872694134712219, + 0.07699986547231674, + 0.04111647978425026, + 0.0700746402144432, + 0.059127841144800186, + 0.057470884174108505, + 0.17306312918663025, + 0.09438320249319077, + 0.09043664485216141, + 0.05005786567926407 + ], + "credit_map": [ + 0.027378451079130173, + 0.02102523110806942, + 0.03701946884393692, + 0.03462442010641098, + 0.05932891368865967, + 0.019608011469244957, + 0.03505001962184906, + 0.053234897553920746, + 0.009118740446865559, + 0.008374550379812717, + 0.010078759863972664, + 0.008309898898005486, + 0.011633320711553097, + 0.0073815626092255116, + 0.012089542113244534, + 0.010013499297201633, + 0.0047902935184538364, + 0.004963215906172991, + 0.004990987945348024, + 0.00484587112441659, + 0.005047551356256008, + 0.004369509406387806, + 0.005024389363825321, + 0.007084662094712257, + 0.008695824071764946, + 0.006618218030780554, + 0.008501702919602394, + 0.00795937329530716, + 0.010099723935127258, + 0.00912303477525711, + 0.009771925397217274, + 0.009304841980338097, + 0.005578149110078812, + 0.009254587814211845, + 0.007491781376302242, + 0.0060641951858997345, + 0.009258824400603771, + 0.005803444888442755, + 0.010898503474891186, + 0.004778355360031128, + 0.003641443094238639, + 0.003696705447509885, + 0.007729634642601013, + 0.004483977798372507, + 0.020204344764351845, + 0.006816011853516102, + 0.007675475906580687, + 0.003223288571462035, + 0.005765700712800026, + 0.018570831045508385, + 0.008236095309257507, + 0.040183354169130325, + 0.06691838055849075, + 0.017486102879047394, + 0.008965364657342434, + 0.006937292870134115, + 0.007426050025969744, + 0.010425345040857792, + 0.012704874388873577, + 0.015315523371100426, + 0.013347923755645752, + 0.012964220717549324, + 0.011457042768597603, + 0.010742232203483582, + 0.006333133205771446, + 0.009879210963845253, + 0.019155096262693405, + 0.008170673623681068, + 0.014313059858977795, + 0.01421490777283907, + 0.011312633752822876, + 0.0070579275488853455, + 0.007144152652472258, + 0.007063194643706083, + 0.006159037351608276, + 0.006829670164734125, + 0.006079546641558409, + 0.006484347861260176, + 0.006063736043870449, + 0.0042341831140220165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.778639018535614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373705595731735 + ], + "opposite_logprob_tokens": [ + -3.4087369441986084 + ], + "logprob_reward": -0.03373705595731735, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373705595731735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3364274501800537, + 0.038223762065172195, + 0.07887773960828781, + 0.07730688899755478, + 0.05303237959742546, + 0.0672636404633522, + 0.2233247607946396, + 0.057892732322216034, + 0.023984204977750778, + 0.04366647079586983 + ], + "credit_map": [ + 0.041357431560754776, + 0.03104162961244583, + 0.04863826557993889, + 0.03829779475927353, + 0.04937838390469551, + 0.03203992918133736, + 0.05701539292931557, + 0.03865859657526016, + 0.005880727898329496, + 0.004184115678071976, + 0.004304461646825075, + 0.004646176937967539, + 0.003867241321131587, + 0.0045319814234972, + 0.004690921399742365, + 0.006118132267147303, + 0.019657693803310394, + 0.008263683877885342, + 0.01182484533637762, + 0.005019410979002714, + 0.006619737483561039, + 0.009302977472543716, + 0.00871995184570551, + 0.009469435550272465, + 0.005835886113345623, + 0.009975709952414036, + 0.018089547753334045, + 0.007883141748607159, + 0.01365957036614418, + 0.005151628982275724, + 0.011004353873431683, + 0.005707046017050743, + 0.005182297434657812, + 0.008900085464119911, + 0.006428271997720003, + 0.005427572876214981, + 0.007066748570650816, + 0.006225181743502617, + 0.007637493312358856, + 0.006164726801216602, + 0.004082520958036184, + 0.0041770110838115215, + 0.010446137748658657, + 0.005438121501356363, + 0.01928088068962097, + 0.009860605001449585, + 0.008468050509691238, + 0.005510305520147085, + 0.006925382185727358, + 0.02314576506614685, + 0.0095017459243536, + 0.05332191288471222, + 0.09224796295166016, + 0.022272400557994843, + 0.009296415373682976, + 0.006613172125071287, + 0.006266779266297817, + 0.008765234611928463, + 0.006893626414239407, + 0.009085076861083508, + 0.005990250036120415, + 0.006823228672146797, + 0.006103062070906162, + 0.007965469732880592, + 0.0028333805967122316, + 0.0030919997952878475, + 0.0034735258668661118, + 0.002903126645833254, + 0.0037818585988134146, + 0.0026660405565053225, + 0.003571458626538515, + 0.0016628146404400468, + 0.0046892897225916386, + 0.006390954367816448, + 0.005590875633060932, + 0.006744348909705877, + 0.00651037460193038, + 0.005725027061998844, + 0.004813680425286293, + 0.0032019210048019886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.03373705595731735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019456282025203109 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019456282025203109, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019456282025203109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17064131796360016, + 0.2388850748538971, + 0.03962966799736023, + 0.05370926856994629, + 0.040718477219343185, + 0.058586083352565765, + 0.20233649015426636, + 0.045633863657712936, + 0.09945333749055862, + 0.05040645971894264 + ], + "credit_map": [ + 0.015017279423773289, + 0.012469654902815819, + 0.020725391805171967, + 0.015411430969834328, + 0.04045522212982178, + 0.012545564211905003, + 0.0224026869982481, + 0.03161408007144928, + 0.01803441531956196, + 0.03431927412748337, + 0.06179223209619522, + 0.029163740575313568, + 0.027029849588871002, + 0.016966473311185837, + 0.027332352474331856, + 0.02424672059714794, + 0.003966009709984064, + 0.005175471771508455, + 0.0060998606495559216, + 0.00451697688549757, + 0.008099484257400036, + 0.003778645070269704, + 0.003803927218541503, + 0.004189294297248125, + 0.006685263942927122, + 0.006250648759305477, + 0.00916789285838604, + 0.005167015362530947, + 0.008229314349591732, + 0.007250653114169836, + 0.006230071187019348, + 0.004728403873741627, + 0.003673162776976824, + 0.00510351313278079, + 0.005866459105163813, + 0.004256364423781633, + 0.007725093048065901, + 0.003318427363410592, + 0.006615839432924986, + 0.004159616772085428, + 0.004545337986201048, + 0.003928124904632568, + 0.007518396247178316, + 0.004861123394221067, + 0.014960769563913345, + 0.010098813101649284, + 0.007850325666368008, + 0.0048231869004666805, + 0.006880360189825296, + 0.023083718493580818, + 0.008665894158184528, + 0.044902071356773376, + 0.08280277252197266, + 0.0209719967097044, + 0.008865335024893284, + 0.006164324935525656, + 0.007463792804628611, + 0.007763344328850508, + 0.005230346694588661, + 0.0036124163307249546, + 0.007814774289727211, + 0.005945035722106695, + 0.003510416252538562, + 0.004293737933039665, + 0.010573162697255611, + 0.012968757189810276, + 0.017910437658429146, + 0.009853585623204708, + 0.01724434643983841, + 0.010957772843539715, + 0.012764287181198597, + 0.007180979009717703, + 0.007202502340078354, + 0.009264715015888214, + 0.004554979503154755, + 0.006938927341252565, + 0.00583122530952096, + 0.006152385380119085, + 0.006878457963466644, + 0.00358326337300241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0019456282025203109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001519954763352871 + ], + "opposite_logprob_tokens": [ + -6.501520156860352 + ], + "logprob_reward": -0.001519954763352871, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001519954763352871, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14373980462551117, + 0.0548354797065258, + 0.05445773899555206, + 0.032535191625356674, + 0.06868720799684525, + 0.06854365020990372, + 0.30356934666633606, + 0.13373589515686035, + 0.05568673834204674, + 0.0842089131474495 + ], + "credit_map": [ + 0.01189678255468607, + 0.00855250284075737, + 0.012719978578388691, + 0.013037738390266895, + 0.03702481463551521, + 0.00954772625118494, + 0.01837220788002014, + 0.03258807212114334, + 0.005093249026685953, + 0.006513675209134817, + 0.010585566982626915, + 0.006772590801119804, + 0.008471694774925709, + 0.004772192798554897, + 0.006321770139038563, + 0.006304747890681028, + 0.007689714431762695, + 0.007509126327931881, + 0.0063394163735210896, + 0.0038611062336713076, + 0.011632176116108894, + 0.006170189473778009, + 0.006156955845654011, + 0.005099060013890266, + 0.0032082258258014917, + 0.0031733892392367125, + 0.005644408520311117, + 0.0031702457927167416, + 0.004879991058260202, + 0.00393723463639617, + 0.0052657341584563255, + 0.003255964256823063, + 0.0063367486000061035, + 0.013631360605359077, + 0.007961356081068516, + 0.006515725050121546, + 0.011562398634850979, + 0.008158487267792225, + 0.008005376905202866, + 0.00651576416566968, + 0.003084843745455146, + 0.00367542402818799, + 0.008204747922718525, + 0.0043454780243337154, + 0.02457137405872345, + 0.011800076812505722, + 0.008019516244530678, + 0.0048421937972307205, + 0.008928066119551659, + 0.03179521858692169, + 0.01352792326360941, + 0.07371493428945541, + 0.12473703920841217, + 0.03015381656587124, + 0.012498486787080765, + 0.00821386743336916, + 0.012753942981362343, + 0.01885986514389515, + 0.017587296664714813, + 0.02257925271987915, + 0.02248654142022133, + 0.018199287354946136, + 0.011286315508186817, + 0.009983418509364128, + 0.004801800474524498, + 0.006834586150944233, + 0.012264197692275047, + 0.00617220439016819, + 0.007784866262227297, + 0.009772446006536484, + 0.004862230271100998, + 0.003194410353899002, + 0.009291941300034523, + 0.013915365561842918, + 0.00787833146750927, + 0.011674648150801659, + 0.013055897317826748, + 0.008987926878035069, + 0.01281613577157259, + 0.0065886774100363255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.001519954763352871 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028183048125356436 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.0028183048125356436, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028183048125356436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11669804155826569, + 0.04943527653813362, + 0.036978110671043396, + 0.03607059270143509, + 0.051822301000356674, + 0.06272441893815994, + 0.3057316839694977, + 0.08209194242954254, + 0.08398576825857162, + 0.17446185648441315 + ], + "credit_map": [ + 0.007804121822118759, + 0.009850708767771721, + 0.014248614199459553, + 0.011527995578944683, + 0.03360573574900627, + 0.006496488582342863, + 0.01552396360784769, + 0.017640413716435432, + 0.005901044234633446, + 0.005110627040266991, + 0.010886065661907196, + 0.005002778489142656, + 0.007007249165326357, + 0.0045838612131774426, + 0.005863916128873825, + 0.005079737398773432, + 0.00432231929153204, + 0.005692502483725548, + 0.004211952909827232, + 0.0036906383465975523, + 0.007590743247419596, + 0.0035032564774155617, + 0.00389100331813097, + 0.004075694363564253, + 0.0031924094073474407, + 0.004205901175737381, + 0.008723385632038116, + 0.003381487214937806, + 0.005691836588084698, + 0.003194347256794572, + 0.004578637424856424, + 0.00310258730314672, + 0.004697361495345831, + 0.005664757918566465, + 0.008498236536979675, + 0.006840004585683346, + 0.008774293586611748, + 0.005422572139650583, + 0.00771922105923295, + 0.004205856937915087, + 0.004055303521454334, + 0.0034722108393907547, + 0.007841451093554497, + 0.004031675402075052, + 0.020066913217306137, + 0.009159916080534458, + 0.008350114338099957, + 0.005746831651777029, + 0.00912458449602127, + 0.033281199634075165, + 0.012421844527125359, + 0.07322898507118225, + 0.1266905516386032, + 0.02996881492435932, + 0.012842702679336071, + 0.008172999136149883, + 0.01434167567640543, + 0.011106072925031185, + 0.010145938955247402, + 0.011460917070508003, + 0.012077104300260544, + 0.010576123371720314, + 0.0066390917636454105, + 0.0057450151070952415, + 0.0090255131945014, + 0.011505452916026115, + 0.01160883717238903, + 0.00832823570817709, + 0.014736102893948555, + 0.007800978608429432, + 0.013898788020014763, + 0.007081859279423952, + 0.01801246777176857, + 0.028956232592463493, + 0.014556124806404114, + 0.020235700532794, + 0.025745080783963203, + 0.018756641075015068, + 0.027817554771900177, + 0.02038205787539482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0028183048125356436 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long, overhead shot, slowly tilting upward from street level, plunges the viewer into a cyberpunk future drenched in perpetual acid rain and shimmering holographic advertisements. Fine acid rain falls like a curtain, creating tiny ripples on the slippery asphalt, while massive, volatile neon advertisements projected between the glass facades of skyscrapers cast eerie, hypersaturated blues, purples, and pinks against the gloomy night sky. Against this chaotic backdrop, the colossal Godzilla clashes fiercely with a comparable mechanical behemoth. Godzilla's rough, rock-gray scales gleam wetly in the rain, an ominous atomic blue shimmering through the gaps between them. Godzilla leans forward slightly, muscles tense, brimming with primal fury. The opposing behemoth is constructed of polished titanium and carbon fiber, its geometric armor gleaming with a cold metallic sheen. Complex hydraulic lines and glowing red sensors crisscross its body, its shoulder-mounted weapon systems activated. Suddenly, Godzilla's dorsal fins began to light up one by one, emitting a pulsating blue light, and a hot atomic breath suddenly burst out from its open jaws. At the same time, the core reactor in the chest of the mechanical monster burst into a dazzling white light, and a precise beam of plasma energy shot towards it. The two destructive energies collided violently in mid-air, producing a spherical, distorted energy explosion, which instantly split the gloomy sky and illuminated the entire city, creating a tense, grand and doomsday atmosphere of epic confrontation.", + "seed": 15458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15458_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13063876887317746, + "vlm_logprob": -0.13063876887317746, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.13063876887317746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.597930371761322 + ], + "opposite_logprob_tokens": [ + -0.847930371761322 + ], + "logprob_reward": -0.597930371761322, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.597930371761322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11448501795530319, + 0.07174480706453323, + 0.10566720366477966, + 0.08366098999977112, + 0.11768791079521179, + 0.044021714478731155, + 0.23666688799858093, + 0.08490709960460663, + 0.044154565781354904, + 0.09700378775596619 + ], + "credit_map": [ + 0.010914456099271774, + 0.011793488636612892, + 0.018745217472314835, + 0.014335082843899727, + 0.011058073490858078, + 0.015038582496345043, + 0.013959537260234356, + 0.018640585243701935, + 0.006062151398509741, + 0.005304805003106594, + 0.01847056858241558, + 0.00810926128178835, + 0.01005815714597702, + 0.007109780330210924, + 0.0103994095697999, + 0.00623067282140255, + 0.009656116366386414, + 0.010394453071057796, + 0.008856465108692646, + 0.008570589125156403, + 0.018781453371047974, + 0.028047122061252594, + 0.008305628784000874, + 0.013055364601314068, + 0.00831367913633585, + 0.008462964557111263, + 0.013013757765293121, + 0.011359850876033306, + 0.009870667941868305, + 0.009865638799965382, + 0.014277839101850986, + 0.008496589958667755, + 0.011562718078494072, + 0.009676854126155376, + 0.010909532196819782, + 0.007436895743012428, + 0.03664674609899521, + 0.007852822542190552, + 0.025771822780370712, + 0.007830513641238213, + 0.003681328147649765, + 0.004522399045526981, + 0.005406952928751707, + 0.005235811695456505, + 0.009249802678823471, + 0.0038850249256938696, + 0.008815661072731018, + 0.0032247332856059074, + 0.010816081427037716, + 0.013006014749407768, + 0.04123879224061966, + 0.013226422481238842, + 0.10087677836418152, + 0.016536004841327667, + 0.030247775837779045, + 0.010719028301537037, + 0.007412333972752094, + 0.0083556417375803, + 0.014466123655438423, + 0.012085323221981525, + 0.01303176861256361, + 0.011077236384153366, + 0.011618795804679394, + 0.006859876681119204, + 0.003580337157472968, + 0.007972068153321743, + 0.0031039321329444647, + 0.005635538604110479, + 0.007586840074509382, + 0.0044789304956793785, + 0.006803188938647509, + 0.004993727896362543, + 0.009367107413709164, + 0.01007704995572567, + 0.010001817718148232, + 0.010190554894506931, + 0.01805201731622219, + 0.011614570394158363, + 0.014700476080179214, + 0.013000192120671272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.597930371761322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868645966053009 + ], + "opposite_logprob_tokens": [ + -3.0486865043640137 + ], + "logprob_reward": -0.04868645966053009, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868645966053009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22657649219036102, + 0.03521445393562317, + 0.06773105263710022, + 0.050387658178806305, + 0.06737680733203888, + 0.06979589909315109, + 0.28339484333992004, + 0.03477040305733681, + 0.024591507390141487, + 0.14016084372997284 + ], + "credit_map": [ + 0.02095889300107956, + 0.027051959186792374, + 0.02341596782207489, + 0.041835423558950424, + 0.024518506601452827, + 0.020096395164728165, + 0.04406066611409187, + 0.024638716131448746, + 0.004294678568840027, + 0.00472001638263464, + 0.004247266333550215, + 0.0037988030817359686, + 0.006796001922339201, + 0.003625999903306365, + 0.003704256610944867, + 0.004027432296425104, + 0.0047874776646494865, + 0.005376617424190044, + 0.010613769292831421, + 0.016481619328260422, + 0.005366839934140444, + 0.01210800465196371, + 0.004343019332736731, + 0.008653710596263409, + 0.004953280556946993, + 0.00662369467318058, + 0.006301016081124544, + 0.006704165134578943, + 0.006909795105457306, + 0.0036693881265819073, + 0.010391502641141415, + 0.004834817722439766, + 0.004246105439960957, + 0.003983667120337486, + 0.007335910573601723, + 0.004670274443924427, + 0.023852208629250526, + 0.004320917185395956, + 0.015590088441967964, + 0.0033776445779949427, + 0.0059443372301757336, + 0.006579260341823101, + 0.006126823835074902, + 0.005952951032668352, + 0.01322114933282137, + 0.005900158081203699, + 0.021165484562516212, + 0.004905741196125746, + 0.012832190841436386, + 0.01353274192661047, + 0.0550181120634079, + 0.014626006595790386, + 0.12717002630233765, + 0.019938865676522255, + 0.03216734156012535, + 0.008109573274850845, + 0.003093692706897855, + 0.0032874613534659147, + 0.0027342697139829397, + 0.007159633561968803, + 0.0038766206707805395, + 0.004277766216546297, + 0.006421081256121397, + 0.003919882699847221, + 0.003151845885440707, + 0.0034709391184151173, + 0.0034572433214634657, + 0.0037398014683276415, + 0.0027549853548407555, + 0.0027826556470245123, + 0.002996725495904684, + 0.00223731342703104, + 0.014791704714298248, + 0.01455837395042181, + 0.01594298519194126, + 0.014032263308763504, + 0.024078134447336197, + 0.021088123321533203, + 0.01973782107234001, + 0.015931453555822372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04868645966053009 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031956578604876995 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.0031956578604876995, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031956578604876995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0601576566696167, + 0.03713926672935486, + 0.06762801855802536, + 0.16660776734352112, + 0.11897902190685272, + 0.0938837006688118, + 0.22097229957580566, + 0.12970802187919617, + 0.04381095618009567, + 0.061113324016332626 + ], + "credit_map": [ + 0.006608641240745783, + 0.007305349223315716, + 0.007052879314869642, + 0.008335231803357601, + 0.006618686951696873, + 0.007795892655849457, + 0.008252904750406742, + 0.008188074454665184, + 0.003239659359678626, + 0.0028630841989070177, + 0.005491324234753847, + 0.004429742693901062, + 0.005702702794224024, + 0.003862682031467557, + 0.00867136474698782, + 0.002878705970942974, + 0.005469268653541803, + 0.012330224737524986, + 0.006438924930989742, + 0.007327151019126177, + 0.011171253398060799, + 0.010686437599360943, + 0.006916028447449207, + 0.007288730703294277, + 0.015450247563421726, + 0.016017377376556396, + 0.03563883528113365, + 0.017901666462421417, + 0.026080967858433723, + 0.01038275845348835, + 0.027364583685994148, + 0.01777133159339428, + 0.015301923267543316, + 0.009504851885139942, + 0.016416989266872406, + 0.013779260218143463, + 0.01766168139874935, + 0.012350905686616898, + 0.02143423818051815, + 0.01252917293459177, + 0.00969515647739172, + 0.009227368980646133, + 0.007132640108466148, + 0.005940948147326708, + 0.01776735484600067, + 0.009103653021156788, + 0.028553517535328865, + 0.006463063880801201, + 0.009872524067759514, + 0.010515769943594933, + 0.04470568895339966, + 0.011601313017308712, + 0.09725522994995117, + 0.016609834507107735, + 0.024203287437558174, + 0.0062086512334644794, + 0.008838978596031666, + 0.01185406744480133, + 0.01412664819508791, + 0.012187673710286617, + 0.024462729692459106, + 0.0187559612095356, + 0.030183324590325356, + 0.009298635646700859, + 0.004542500711977482, + 0.007172165904194117, + 0.0026188206393271685, + 0.006438526790589094, + 0.005548201035708189, + 0.005508145783096552, + 0.0071358345448970795, + 0.0048467605374753475, + 0.006919265724718571, + 0.0071557071059942245, + 0.006613005418330431, + 0.006438591983169317, + 0.010899106040596962, + 0.007840856909751892, + 0.008598648943006992, + 0.006648141425102949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.0031956578604876995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011851914459839463 + ], + "opposite_logprob_tokens": [ + -6.751185417175293 + ], + "logprob_reward": -0.0011851914459839463, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011851914459839463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18448932468891144, + 0.0769323781132698, + 0.10782475769519806, + 0.154582679271698, + 0.035779453814029694, + 0.0517524778842926, + 0.10526733100414276, + 0.07513127475976944, + 0.1478375941514969, + 0.06040272116661072 + ], + "credit_map": [ + 0.015009847469627857, + 0.03104235790669918, + 0.025739049538969994, + 0.03319982811808586, + 0.017972249537706375, + 0.02055950090289116, + 0.02865295670926571, + 0.012313530780375004, + 0.007079857401549816, + 0.006829575635492802, + 0.01653211936354637, + 0.0070137036964297295, + 0.013078050687909126, + 0.008114536292850971, + 0.011400632560253143, + 0.006883902009576559, + 0.009387887082993984, + 0.009232338517904282, + 0.01459695864468813, + 0.021808680146932602, + 0.012163685634732246, + 0.02265322022140026, + 0.008183392696082592, + 0.009798598475754261, + 0.016461828723549843, + 0.01324599701911211, + 0.04584123566746712, + 0.02042873203754425, + 0.013819132000207901, + 0.0084233358502388, + 0.013870032504200935, + 0.022492386400699615, + 0.003927886951714754, + 0.004174028057605028, + 0.004680763464421034, + 0.004931955132633448, + 0.004847937263548374, + 0.003980105742812157, + 0.005980047397315502, + 0.003256730269640684, + 0.003606934566050768, + 0.005243409890681505, + 0.007378957699984312, + 0.006244869437068701, + 0.013239521533250809, + 0.004213384818285704, + 0.008933357894420624, + 0.002892040880396962, + 0.00540360901504755, + 0.006814952939748764, + 0.017780335620045662, + 0.006557651795446873, + 0.03470954671502113, + 0.014901753515005112, + 0.010396036319434643, + 0.008703445084393024, + 0.0064554596319794655, + 0.00781219033524394, + 0.007689363323152065, + 0.009574721567332745, + 0.008088774047791958, + 0.009047092869877815, + 0.016412077471613884, + 0.010051597841084003, + 0.016431020572781563, + 0.02022029459476471, + 0.018497535958886147, + 0.023607630282640457, + 0.02005716785788536, + 0.014432569965720177, + 0.02041127160191536, + 0.014180107042193413, + 0.006861418019980192, + 0.00652362871915102, + 0.005665226839482784, + 0.00573338195681572, + 0.009769260883331299, + 0.0117954695597291, + 0.007122252136468887, + 0.006932082120329142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0011851914459839463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021961636375635862 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021961636375635862, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021961636375635862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08456870913505554, + 0.07015269249677658, + 0.06488864123821259, + 0.08967428654432297, + 0.0773499459028244, + 0.06096695736050606, + 0.19711142778396606, + 0.11313586682081223, + 0.12707217037677765, + 0.11507932096719742 + ], + "credit_map": [ + 0.006933592725545168, + 0.010519573464989662, + 0.015160943381488323, + 0.011704833246767521, + 0.008831973187625408, + 0.011173227801918983, + 0.011780942790210247, + 0.008463605307042599, + 0.006597147323191166, + 0.005269533488899469, + 0.014788745902478695, + 0.007459952030330896, + 0.01143107283860445, + 0.007195807062089443, + 0.008740933611989021, + 0.008669495582580566, + 0.005865687504410744, + 0.010608445852994919, + 0.008040091954171658, + 0.013222184032201767, + 0.006112202536314726, + 0.008498039096593857, + 0.0068184686824679375, + 0.005723514128476381, + 0.008668051101267338, + 0.008552362211048603, + 0.022554831579327583, + 0.010435285046696663, + 0.00929243117570877, + 0.007288456428796053, + 0.010636146180331707, + 0.01224671583622694, + 0.005171958357095718, + 0.004566112067550421, + 0.0074053010903298855, + 0.006689806003123522, + 0.026782343164086342, + 0.003973645158112049, + 0.01786772720515728, + 0.004893038887530565, + 0.0057503459975123405, + 0.007016715593636036, + 0.006790814455598593, + 0.00524086644873023, + 0.01342417486011982, + 0.005696380976587534, + 0.013496502302587032, + 0.0035511532332748175, + 0.008828462101519108, + 0.010048411786556244, + 0.041045866906642914, + 0.010093467310070992, + 0.08648911118507385, + 0.01336223166435957, + 0.02117127925157547, + 0.006072572898119688, + 0.0074661727994680405, + 0.010683879256248474, + 0.009695352055132389, + 0.021449895575642586, + 0.016956539824604988, + 0.013899620622396469, + 0.024393849074840546, + 0.008590543642640114, + 0.010820678435266018, + 0.017662908881902695, + 0.011512592434883118, + 0.019167713820934296, + 0.020284684374928474, + 0.012377766892313957, + 0.016219334676861763, + 0.019026482477784157, + 0.014739100821316242, + 0.012695795856416225, + 0.012051797471940517, + 0.01097016129642725, + 0.02365201897919178, + 0.011715123429894447, + 0.015408096835017204, + 0.01384721975773573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.0021961636375635862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long, overhead shot, slowly tilting upward from street level, plunges the viewer into a cyberpunk future drenched in perpetual acid rain and shimmering holographic advertisements. Fine acid rain falls like a curtain, creating tiny ripples on the slippery asphalt, while massive, volatile neon advertisements projected between the glass facades of skyscrapers cast eerie, hypersaturated blues, purples, and pinks against the gloomy night sky. Against this chaotic backdrop, the colossal Godzilla clashes fiercely with a comparable mechanical behemoth. Godzilla's rough, rock-gray scales gleam wetly in the rain, an ominous atomic blue shimmering through the gaps between them. Godzilla leans forward slightly, muscles tense, brimming with primal fury. The opposing behemoth is constructed of polished titanium and carbon fiber, its geometric armor gleaming with a cold metallic sheen. Complex hydraulic lines and glowing red sensors crisscross its body, its shoulder-mounted weapon systems activated. Suddenly, Godzilla's dorsal fins began to light up one by one, emitting a pulsating blue light, and a hot atomic breath suddenly burst out from its open jaws. At the same time, the core reactor in the chest of the mechanical monster burst into a dazzling white light, and a precise beam of plasma energy shot towards it. The two destructive energies collided violently in mid-air, producing a spherical, distorted energy explosion, which instantly split the gloomy sky and illuminated the entire city, creating a tense, grand and doomsday atmosphere of epic confrontation.", + "seed": 15459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_3/index_4208_seed_15459_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19580081193707882, + "vlm_logprob": -0.19580081193707882, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.19580081193707882, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.854393482208252 + ], + "opposite_logprob_tokens": [ + -0.604393482208252 + ], + "logprob_reward": -0.854393482208252, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.854393482208252, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4845113158226013, + 0.06906047463417053, + 0.029750581830739975, + 0.027519473806023598, + 0.045540519058704376, + 0.04229922220110893, + 0.05895601958036423, + 0.09691936522722244, + 0.0763559564948082, + 0.06908702850341797 + ], + "credit_map": [ + 0.054787565022706985, + 0.045330487191677094, + 0.0482519194483757, + 0.04584473744034767, + 0.11245803534984589, + 0.07567562907934189, + 0.05501227080821991, + 0.04715067520737648, + 0.0058966102078557014, + 0.007336112204939127, + 0.00833465252071619, + 0.009660872630774975, + 0.012083747424185276, + 0.009766848757863045, + 0.008893615566194057, + 0.007088017649948597, + 0.002249648328870535, + 0.0031499071046710014, + 0.004013144411146641, + 0.0035576396621763706, + 0.0070616514421999454, + 0.004079543985426426, + 0.003201547311618924, + 0.0024374991189688444, + 0.0024979470763355494, + 0.003017012495547533, + 0.0034271329641342163, + 0.003472487209364772, + 0.0046598161570727825, + 0.003035033820196986, + 0.004867312964051962, + 0.0025427311193197966, + 0.0032202929724007845, + 0.003923372831195593, + 0.007891004905104637, + 0.0035591202322393656, + 0.006385826505720615, + 0.004891387186944485, + 0.012184812687337399, + 0.0034846970811486244, + 0.0033247899264097214, + 0.005723074544221163, + 0.005923018790781498, + 0.004520397167652845, + 0.006416788790374994, + 0.0038926093839108944, + 0.005082949995994568, + 0.0074155922047793865, + 0.004707639571279287, + 0.011912052519619465, + 0.007250553462654352, + 0.011288397014141083, + 0.008025359362363815, + 0.004788761027157307, + 0.004381422419101, + 0.006601838394999504, + 0.018120499327778816, + 0.010917287319898605, + 0.011584683321416378, + 0.009990139864385128, + 0.014533757232129574, + 0.005119249690324068, + 0.011703851632773876, + 0.014949900098145008, + 0.00978002231568098, + 0.011032001115381718, + 0.008457429707050323, + 0.02112141251564026, + 0.005105006508529186, + 0.0052073379047214985, + 0.008211162872612476, + 0.007441583089530468, + 0.019438982009887695, + 0.013234487734735012, + 0.003341873176395893, + 0.004195154178887606, + 0.003895331872627139, + 0.004873186815530062, + 0.006671924144029617, + 0.013436089269816875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.854393482208252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299623548984528 + ], + "opposite_logprob_tokens": [ + -2.2379963397979736 + ], + "logprob_reward": -0.11299623548984528, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299623548984528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3948945105075836, + 0.07137136906385422, + 0.04628470540046692, + 0.09923583269119263, + 0.045143868774175644, + 0.055875785648822784, + 0.14325635135173798, + 0.06624935567378998, + 0.042773015797138214, + 0.03491514176130295 + ], + "credit_map": [ + 0.041511546820402145, + 0.03510674461722374, + 0.039171092212200165, + 0.04499022662639618, + 0.07911733537912369, + 0.06425883620977402, + 0.051018837839365005, + 0.03971990942955017, + 0.006205644458532333, + 0.006585085764527321, + 0.008081485517323017, + 0.009441816247999668, + 0.015080837532877922, + 0.00918575283139944, + 0.009334205649793148, + 0.007456544786691666, + 0.004186186008155346, + 0.006153955589979887, + 0.0055112820118665695, + 0.004686205182224512, + 0.006161329336464405, + 0.00891542062163353, + 0.007324179168790579, + 0.003346146084368229, + 0.009106702171266079, + 0.01176598109304905, + 0.013578255660831928, + 0.012042236514389515, + 0.009714120998978615, + 0.010241921059787273, + 0.025225333869457245, + 0.00756129389628768, + 0.004096501506865025, + 0.004449428990483284, + 0.006110839545726776, + 0.0038986019790172577, + 0.008004296571016312, + 0.003966846037656069, + 0.010141966864466667, + 0.004475386813282967, + 0.003928795922547579, + 0.014218191616237164, + 0.01209657546132803, + 0.004775632172822952, + 0.00627765990793705, + 0.0049747503362596035, + 0.005892940331250429, + 0.003711237572133541, + 0.010067999362945557, + 0.04145906865596771, + 0.016556544229388237, + 0.024124007672071457, + 0.021245311945676804, + 0.009973566979169846, + 0.009549160487949848, + 0.010280700400471687, + 0.007774641737341881, + 0.004824535921216011, + 0.006593667436391115, + 0.005122314672917128, + 0.017711563035845757, + 0.006100142374634743, + 0.003920151386409998, + 0.014202338643372059, + 0.006266032345592976, + 0.005323630291968584, + 0.004041972104460001, + 0.008418770506978035, + 0.0028554857708513737, + 0.003909379243850708, + 0.006029688287526369, + 0.005928054917603731, + 0.0072367931716144085, + 0.005944162607192993, + 0.003052733140066266, + 0.005019811447709799, + 0.002271282020956278, + 0.0024116800632327795, + 0.004397362936288118, + 0.004581316374242306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.11299623548984528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643134772777557 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643134772777557, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643134772777557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38395994901657104, + 0.06573754549026489, + 0.10644266754388809, + 0.0953821912407875, + 0.06085862219333649, + 0.048049818724393845, + 0.0436464361846447, + 0.05989554151892662, + 0.08297199755907059, + 0.053055234253406525 + ], + "credit_map": [ + 0.045559901744127274, + 0.034025054425001144, + 0.040962494909763336, + 0.0371498167514801, + 0.07748589664697647, + 0.055593669414520264, + 0.05043787881731987, + 0.04274524375796318, + 0.004827549215406179, + 0.0060950531624257565, + 0.009645683690905571, + 0.0061724684201180935, + 0.00978308916091919, + 0.0068881576880812645, + 0.011441400274634361, + 0.01088414341211319, + 0.007367272395640612, + 0.010693812742829323, + 0.015377378091216087, + 0.011440165340900421, + 0.023393943905830383, + 0.016770504415035248, + 0.012558829970657825, + 0.00884075928479433, + 0.005592096596956253, + 0.009162423200905323, + 0.01445817295461893, + 0.010287459008395672, + 0.01887841708958149, + 0.01045879814773798, + 0.02072092890739441, + 0.00582389859482646, + 0.00492125703021884, + 0.006839201785624027, + 0.009024023078382015, + 0.004769561812281609, + 0.009086018428206444, + 0.013511729426681995, + 0.007244108244776726, + 0.005462725646793842, + 0.0026958012022078037, + 0.011584073305130005, + 0.01757851615548134, + 0.0030140967573970556, + 0.004036914091557264, + 0.002386427717283368, + 0.0029678591527044773, + 0.003786133835092187, + 0.004062318243086338, + 0.007472685072571039, + 0.004156588576734066, + 0.00958286039531231, + 0.0065498193725943565, + 0.0029221572913229465, + 0.0037778860423713923, + 0.005122119095176458, + 0.01060048583894968, + 0.006404626648873091, + 0.008511921390891075, + 0.006992251146584749, + 0.009355677291750908, + 0.003346972167491913, + 0.004910425748676062, + 0.009773178957402706, + 0.009257584810256958, + 0.009637480601668358, + 0.01202518679201603, + 0.014478259719908237, + 0.0063156550750136375, + 0.007211067248135805, + 0.014078028500080109, + 0.009968734346330166, + 0.012279217131435871, + 0.008405552245676517, + 0.0035397678148001432, + 0.005303548648953438, + 0.0038209762424230576, + 0.004894685000181198, + 0.006041742395609617, + 0.008769744075834751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.004643134772777557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017200212460011244 + ], + "opposite_logprob_tokens": [ + -6.376719951629639 + ], + "logprob_reward": -0.0017200212460011244, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017200212460011244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27510568499565125, + 0.07441797852516174, + 0.03140275180339813, + 0.06514668464660645, + 0.06085740774869919, + 0.10452111810445786, + 0.06343814730644226, + 0.04192443937063217, + 0.1701761782169342, + 0.11300969868898392 + ], + "credit_map": [ + 0.031420085579156876, + 0.028627395629882812, + 0.03123156540095806, + 0.02841331623494625, + 0.05871402472257614, + 0.03658684715628624, + 0.032770782709121704, + 0.027341630309820175, + 0.006494514644145966, + 0.005201353691518307, + 0.008133137598633766, + 0.006802897434681654, + 0.0076082563027739525, + 0.008482379838824272, + 0.013658142648637295, + 0.018037287518382072, + 0.002560249762609601, + 0.002986133098602295, + 0.0037059870082885027, + 0.003816763637587428, + 0.00510037038475275, + 0.005537606310099363, + 0.004329022951424122, + 0.0033666156232357025, + 0.004304020665585995, + 0.004702391102910042, + 0.008879384957253933, + 0.007613944821059704, + 0.009692373685538769, + 0.00695772934705019, + 0.018817072734236717, + 0.004179765470325947, + 0.004212282598018646, + 0.006385098211467266, + 0.011329219676554203, + 0.006483635865151882, + 0.014715727418661118, + 0.004998944699764252, + 0.008081379346549511, + 0.0046511185355484486, + 0.006460942793637514, + 0.0182639230042696, + 0.024678584188222885, + 0.01028688345104456, + 0.012763396836817265, + 0.009942199103534222, + 0.009077940136194229, + 0.013047237880527973, + 0.0038562039844691753, + 0.01285039633512497, + 0.00833013653755188, + 0.01563023403286934, + 0.006536989472806454, + 0.004421908874064684, + 0.004102119244635105, + 0.007710156962275505, + 0.00527208624407649, + 0.005236171185970306, + 0.004117466509342194, + 0.005551618058234453, + 0.005762030836194754, + 0.0028873616829514503, + 0.006957187317311764, + 0.006140515208244324, + 0.02175016887485981, + 0.025228243321180344, + 0.015479523688554764, + 0.062209904193878174, + 0.0077411760576069355, + 0.008245294913649559, + 0.015020648017525673, + 0.014501205645501614, + 0.03458897024393082, + 0.02300921641290188, + 0.005886385217308998, + 0.0053459517657756805, + 0.005566476844251156, + 0.0064553916454315186, + 0.009653155691921711, + 0.022504139691591263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0017200212460011244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward from street level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a cyberpunk city with perpetual acid rain and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are Godzilla and a mechanical behemoth visible clashing in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla emit a pulsating blue light from its dorsal fins and a burst of atomic breath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mechanical behemoth emit a white core reactor light and a plasma beam that collides with Godzilla's energy?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251185968518257 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251185968518257, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251185968518257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3324832320213318, + 0.04994886741042137, + 0.03633606806397438, + 0.038018546998500824, + 0.0870007574558258, + 0.03997901827096939, + 0.12383034080266953, + 0.03882628306746483, + 0.15456140041351318, + 0.09901547431945801 + ], + "credit_map": [ + 0.03736644983291626, + 0.02655087783932686, + 0.03586656600236893, + 0.03707185387611389, + 0.06499751657247543, + 0.0529254749417305, + 0.0339227132499218, + 0.043781764805316925, + 0.004767909646034241, + 0.006121641956269741, + 0.006339662242680788, + 0.005377232097089291, + 0.007782567758113146, + 0.006582762114703655, + 0.0055769975297153, + 0.007400094531476498, + 0.0031654750928282738, + 0.003903414821252227, + 0.003271718043833971, + 0.0034163868986070156, + 0.006045546382665634, + 0.006773921195417643, + 0.006778344511985779, + 0.0029812625143676996, + 0.003070774022489786, + 0.003020019270479679, + 0.006344534922391176, + 0.003916032612323761, + 0.006372313015162945, + 0.003926908131688833, + 0.009067948907613754, + 0.002300017047673464, + 0.006278244778513908, + 0.005581941921263933, + 0.01260452251881361, + 0.007537537720054388, + 0.019832680001854897, + 0.007285331841558218, + 0.019974393770098686, + 0.007906110025942326, + 0.00252854498103261, + 0.005883951671421528, + 0.009299167431890965, + 0.00430257385596633, + 0.00485718809068203, + 0.003603018820285797, + 0.004264236893504858, + 0.005240337457507849, + 0.008172081783413887, + 0.03621090576052666, + 0.012738536112010479, + 0.027888348326086998, + 0.016613144427537918, + 0.006267635617405176, + 0.007016327232122421, + 0.008923353627324104, + 0.006232536397874355, + 0.004518287722021341, + 0.006486292462795973, + 0.004931070376187563, + 0.00498826801776886, + 0.0030060403514653444, + 0.0037305448204278946, + 0.0049332426860928535, + 0.016840860247612, + 0.01804744452238083, + 0.02402212843298912, + 0.014567553997039795, + 0.011378596536815166, + 0.012706990353763103, + 0.03237573802471161, + 0.024622082710266113, + 0.02062029205262661, + 0.013279504142701626, + 0.007191445678472519, + 0.006788346450775862, + 0.008502436801791191, + 0.010168754495680332, + 0.013846326619386673, + 0.018618369475007057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005251185968518257 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_tensor([3649, 3649])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_tensor([3649, 3649])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f512e43cc8ec4d1cd0e2e240326004600e9fba5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_tensor([3649, 3649])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "crisp, static, top-down, high-angle shot offers a perfect bird's-eye view of a vibrantly bouncing ball as it rebounds powerfully across a clean, expansive wooden tabletop. The table, crafted from warm, honey-toned oak, polished to a soft satin, reveals its delicate, linear grain under the bright, even, and diffuse overhead lighting. The ball itself is a chaotic swirl of electric blue, neon yellow, and hot pink, fused together into a glossy, high-density polymer captured in sharp, circular highlights. With each rhythmic impact, the ball compresses almost imperceptibly against the hardwood before rebounding into the air, its rapid rotation blurring the colors into dynamic streaks as it traces a series of graceful, diminishing arcs across the frame. A soft, perfectly circular shadow dances directly below; its dark twin contracts to a point upon impact and expands as the ball soars, creating a hypnotic, playful atmosphere—a pure and minimalist study in motion, color, and physics.", + "seed": 31997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_3649_seed_31997_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030573198199272157, + "vlm_logprob": -0.030573198199272157, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.030573198199272157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037380870431661606 + ], + "opposite_logprob_tokens": [ + -3.5373809337615967 + ], + "logprob_reward": -0.037380870431661606, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.037380870431661606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1364327222108841, + 0.06510408222675323, + 0.09118910878896713, + 0.2721228003501892, + 0.09561733901500702, + 0.14650924503803253, + 0.07409156858921051, + 0.04835401475429535, + 0.050179623067379, + 0.020399559289216995 + ], + "credit_map": [ + 0.015170227736234665, + 0.014725188724696636, + 0.01837657205760479, + 0.012678808532655239, + 0.01970997452735901, + 0.016738973557949066, + 0.016359053552150726, + 0.02267390862107277, + 0.007619448937475681, + 0.005003160331398249, + 0.009100126102566719, + 0.006991330534219742, + 0.010301177389919758, + 0.00974030327051878, + 0.006960282567888498, + 0.009388241916894913, + 0.009111973457038403, + 0.008739716373383999, + 0.011326259933412075, + 0.011450380086898804, + 0.009288650006055832, + 0.012347659096121788, + 0.01402480062097311, + 0.014899659901857376, + 0.030004072934389114, + 0.04238937795162201, + 0.039700984954833984, + 0.024241473525762558, + 0.03367552161216736, + 0.03673696890473366, + 0.02690737321972847, + 0.03846698999404907, + 0.010012018494307995, + 0.01116475835442543, + 0.01326829195022583, + 0.008072620257735252, + 0.011021984741091728, + 0.01657298021018505, + 0.0076853204518556595, + 0.017819343134760857, + 0.015825074166059494, + 0.012595497071743011, + 0.042978037148714066, + 0.00902431458234787, + 0.01782492734491825, + 0.03018942102789879, + 0.008594298735260963, + 0.009477650746703148, + 0.009455505758523941, + 0.0030931634828448296, + 0.022258667275309563, + 0.0032122370321303606, + 0.003679224057123065, + 0.019375581294298172, + 0.004121757578104734, + 0.008895417675375938, + 0.006397633347660303, + 0.0027083398308604956, + 0.008774522691965103, + 0.0025203244294971228, + 0.0038168684113770723, + 0.009310581721365452, + 0.003376952139660716, + 0.011448781937360764, + 0.01099554542452097, + 0.004003321751952171, + 0.007459361106157303, + 0.004346044268459082, + 0.007260595913976431, + 0.0065778400748968124, + 0.004290624987334013, + 0.00524628721177578, + 0.0022879159078001976, + 0.0040008313953876495, + 0.003596881404519081, + 0.001962296897545457, + 0.0019379775039851665, + 0.0020484726410359144, + 0.002360232640057802, + 0.0022049485705792904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.037380870431661606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059515379369259 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.02059515379369259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059515379369259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20019452273845673, + 0.11003568768501282, + 0.08243244141340256, + 0.08309978991746902, + 0.07585400342941284, + 0.16386209428310394, + 0.0781279131770134, + 0.08849047869443893, + 0.054067328572273254, + 0.0638357475399971 + ], + "credit_map": [ + 0.02829017862677574, + 0.017510171979665756, + 0.05317844823002815, + 0.011740784160792828, + 0.015188097953796387, + 0.03490974381566048, + 0.010192553512752056, + 0.029184551909565926, + 0.025146782398223877, + 0.008828772231936455, + 0.01561253797262907, + 0.014209636487066746, + 0.012895065359771252, + 0.012619424611330032, + 0.005292605143040419, + 0.015430857427418232, + 0.009185855276882648, + 0.005479173269122839, + 0.018435882404446602, + 0.0073792957700788975, + 0.00934861321002245, + 0.01261143758893013, + 0.011619307100772858, + 0.008372883312404156, + 0.011670544743537903, + 0.012625795789062977, + 0.009219562634825706, + 0.009899025782942772, + 0.010206314735114574, + 0.010587588883936405, + 0.01086594071239233, + 0.008025016635656357, + 0.008175990544259548, + 0.00677466532215476, + 0.02079109475016594, + 0.007139522582292557, + 0.006161753088235855, + 0.013349455781280994, + 0.005468407180160284, + 0.007993107661604881, + 0.015761403366923332, + 0.01066338736563921, + 0.058824341744184494, + 0.007905657403171062, + 0.014405230060219765, + 0.04316704720258713, + 0.005918754264712334, + 0.00721628637984395, + 0.009350539185106754, + 0.006203688681125641, + 0.014774879440665245, + 0.008690493181347847, + 0.006272296886891127, + 0.01100147794932127, + 0.009965063072741032, + 0.011869470588862896, + 0.01273581013083458, + 0.005907197017222643, + 0.01753472164273262, + 0.007641599979251623, + 0.00763985188677907, + 0.01799413189291954, + 0.007397642824798822, + 0.011639523319900036, + 0.007710773032158613, + 0.005128278397023678, + 0.01611541211605072, + 0.003441740060225129, + 0.0050751762464642525, + 0.010934218764305115, + 0.0022407476790249348, + 0.0034209825098514557, + 0.006473757792264223, + 0.015599857084453106, + 0.01103023998439312, + 0.006142256315797567, + 0.005815776996314526, + 0.007097093854099512, + 0.0064819492399692535, + 0.005194815341383219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02059515379369259 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866613447666168 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.04866613447666168, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866613447666168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07330606132745743, + 0.05467376857995987, + 0.07804737240076065, + 0.1204315796494484, + 0.16145728528499603, + 0.2809380292892456, + 0.029405726119875908, + 0.09491650760173798, + 0.02762530744075775, + 0.07919841259717941 + ], + "credit_map": [ + 0.009946738369762897, + 0.009066703729331493, + 0.009143832139670849, + 0.007153587881475687, + 0.014302391558885574, + 0.007595074363052845, + 0.007657691836357117, + 0.008440041914582253, + 0.0062547652050852776, + 0.005241211503744125, + 0.010003190487623215, + 0.005092823412269354, + 0.010555887594819069, + 0.008029086515307426, + 0.004932602401822805, + 0.004564205650240183, + 0.007831555791199207, + 0.005545997992157936, + 0.011726905591785908, + 0.00495543796569109, + 0.00906422920525074, + 0.015328237786889076, + 0.0073758563958108425, + 0.01621915027499199, + 0.013066256418824196, + 0.018658677116036415, + 0.02127198502421379, + 0.011044949293136597, + 0.015685422345995903, + 0.017092106863856316, + 0.009272788651287556, + 0.014339386485517025, + 0.02343294955790043, + 0.018909037113189697, + 0.02175038494169712, + 0.01497347466647625, + 0.018670381978154182, + 0.022740241140127182, + 0.012942200526595116, + 0.02803860791027546, + 0.025174209848046303, + 0.01652720756828785, + 0.09730251133441925, + 0.018692724406719208, + 0.019956255331635475, + 0.0682162269949913, + 0.01390847284346819, + 0.021160438656806946, + 0.004865933209657669, + 0.002343543106690049, + 0.006889064330607653, + 0.002167855389416218, + 0.0018918904243037105, + 0.005401544738560915, + 0.00246315635740757, + 0.003382738446816802, + 0.009281122125685215, + 0.0062249102629721165, + 0.015640361234545708, + 0.005466612521559, + 0.006931378971785307, + 0.024980243295431137, + 0.00641665980219841, + 0.019975226372480392, + 0.002925468608736992, + 0.0026487489230930805, + 0.007027625106275082, + 0.0022998054046183825, + 0.0037674657069146633, + 0.00531026441603899, + 0.002142011420801282, + 0.0015039168065413833, + 0.007715374696999788, + 0.006742584053426981, + 0.010882332921028137, + 0.006901734508574009, + 0.006230894010514021, + 0.015926368534564972, + 0.007290173787623644, + 0.017508946359157562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.04866613447666168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369118645787239 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369118645787239, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369118645787239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18638667464256287, + 0.044124338775873184, + 0.07831354439258575, + 0.1702418327331543, + 0.05472191795706749, + 0.062042951583862305, + 0.021008437499403954, + 0.08186976611614227, + 0.235129714012146, + 0.06616085022687912 + ], + "credit_map": [ + 0.016539836302399635, + 0.016156580299139023, + 0.052008818835020065, + 0.012115887366235256, + 0.017800794914364815, + 0.03441891074180603, + 0.01401121262460947, + 0.02333463542163372, + 0.005398847162723541, + 0.004150919616222382, + 0.005878920666873455, + 0.006367005407810211, + 0.007575661409646273, + 0.005276685114949942, + 0.004068603273481131, + 0.0054076979868113995, + 0.007937954738736153, + 0.006584980990737677, + 0.015576229430735111, + 0.006673527415841818, + 0.009080247022211552, + 0.012812278233468533, + 0.009382822550833225, + 0.010265500284731388, + 0.01541381236165762, + 0.02478061057627201, + 0.023661362007260323, + 0.018028266727924347, + 0.021382370963692665, + 0.02218911610543728, + 0.024281054735183716, + 0.020505230873823166, + 0.007096692454069853, + 0.007558120880275965, + 0.00837145745754242, + 0.004339680075645447, + 0.007430838420987129, + 0.008934027515351772, + 0.004081811755895615, + 0.006909285672008991, + 0.0063665625639259815, + 0.005791346542537212, + 0.01451787818223238, + 0.006126555614173412, + 0.006934127304702997, + 0.010135642252862453, + 0.006022744346410036, + 0.006148095708340406, + 0.002487433608621359, + 0.0015863014850765467, + 0.003059796988964081, + 0.002898388309404254, + 0.001650093006901443, + 0.003912930842489004, + 0.0030203210189938545, + 0.0023931728210300207, + 0.00803131703287363, + 0.005578164476901293, + 0.0289889145642519, + 0.0032878026831895113, + 0.006316864397376776, + 0.015426944009959698, + 0.005140642635524273, + 0.009099110029637814, + 0.020514991134405136, + 0.0134165920317173, + 0.08709374070167542, + 0.013678436167538166, + 0.013864060863852501, + 0.06492715328931808, + 0.010134750045835972, + 0.011499982327222824, + 0.004589586518704891, + 0.007971036247909069, + 0.008643475361168385, + 0.005920116789638996, + 0.004636898636817932, + 0.011441265232861042, + 0.006199292838573456, + 0.016759179532527924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.03369118645787239 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532645836472511 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532645836472511, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532645836472511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09916084259748459, + 0.06742295622825623, + 0.03730817139148712, + 0.30659106373786926, + 0.1323665976524353, + 0.06970280408859253, + 0.17442351579666138, + 0.04413073882460594, + 0.02973056025803089, + 0.0391627661883831 + ], + "credit_map": [ + 0.01472032256424427, + 0.008222835138440132, + 0.020429765805602074, + 0.008935702033340931, + 0.010010970756411552, + 0.015034894458949566, + 0.00811297632753849, + 0.013693374581634998, + 0.0067439875565469265, + 0.006352170370519161, + 0.010784387588500977, + 0.007602573372423649, + 0.01066134124994278, + 0.008702441118657589, + 0.006545495707541704, + 0.010030558332800865, + 0.005422474350780249, + 0.0030457754619419575, + 0.006257955450564623, + 0.004005976486951113, + 0.002873990684747696, + 0.006675928831100464, + 0.0038996166549623013, + 0.005126455333083868, + 0.030174655839800835, + 0.03408282622694969, + 0.04965386539697647, + 0.03069235011935234, + 0.02848147787153721, + 0.0491257943212986, + 0.04280877113342285, + 0.04157133400440216, + 0.013841059058904648, + 0.010459283366799355, + 0.03515848144888878, + 0.012425912544131279, + 0.009886233136057854, + 0.025504307821393013, + 0.012822394259274006, + 0.01226892415434122, + 0.006717548705637455, + 0.0038134700153023005, + 0.024291040375828743, + 0.004743883386254311, + 0.003931879997253418, + 0.016820328310132027, + 0.0035663789603859186, + 0.005818274803459644, + 0.01557588204741478, + 0.009537293575704098, + 0.052721407264471054, + 0.013066215440630913, + 0.00940218847244978, + 0.041890714317560196, + 0.015044047497212887, + 0.017185766249895096, + 0.004127723164856434, + 0.004711306653916836, + 0.0147502850741148, + 0.0030278703197836876, + 0.00524981738999486, + 0.0048988936468958855, + 0.003894014284014702, + 0.0034708292223513126, + 0.0022531456779688597, + 0.0037452643737196922, + 0.006763876415789127, + 0.0030780492816120386, + 0.0024600608740001917, + 0.0065900967456400394, + 0.0016076280735433102, + 0.0032324367202818394, + 0.002899289596825838, + 0.009009878151118755, + 0.008605092763900757, + 0.003739125793799758, + 0.0034327171742916107, + 0.003905214834958315, + 0.0045638359151780605, + 0.003007609397172928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.012532645836472511 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crisp, static, top-down, high-angle shot offers a perfect bird's-eye view of a vibrantly bouncing ball as it rebounds powerfully across a clean, expansive wooden tabletop. The table, crafted from warm, honey-toned oak, polished to a soft satin, reveals its delicate, linear grain under the bright, even, and diffuse overhead lighting. The ball itself is a chaotic swirl of electric blue, neon yellow, and hot pink, fused together into a glossy, high-density polymer captured in sharp, circular highlights. With each rhythmic impact, the ball compresses almost imperceptibly against the hardwood before rebounding into the air, its rapid rotation blurring the colors into dynamic streaks as it traces a series of graceful, diminishing arcs across the frame. A soft, perfectly circular shadow dances directly below; its dark twin contracts to a point upon impact and expands as the ball soars, creating a hypnotic, playful atmosphere—a pure and minimalist study in motion, color, and physics.", + "seed": 31998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_3649_seed_31998_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033046736009418966, + "vlm_logprob": -0.033046736009418966, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.033046736009418966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013282505795359612 + ], + "opposite_logprob_tokens": [ + -4.763282299041748 + ], + "logprob_reward": -0.013282505795359612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013282505795359612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10899382829666138, + 0.028213096782565117, + 0.1343241184949875, + 0.04309041053056717, + 0.03528675064444542, + 0.09505290538072586, + 0.046666111797094345, + 0.036751892417669296, + 0.40747973322868347, + 0.06414113193750381 + ], + "credit_map": [ + 0.012534753419458866, + 0.007025605533272028, + 0.01658644899725914, + 0.0178011953830719, + 0.012168909423053265, + 0.017059406265616417, + 0.010085519403219223, + 0.015731994062662125, + 0.003143373178318143, + 0.002688406268134713, + 0.00410107895731926, + 0.003689405508339405, + 0.0028719513211399317, + 0.00463938107714057, + 0.0031918969471007586, + 0.00388760375790298, + 0.017492415383458138, + 0.008081800304353237, + 0.024968966841697693, + 0.017031505703926086, + 0.011848962865769863, + 0.020057229325175285, + 0.012984762899577618, + 0.02185848169028759, + 0.004189853556454182, + 0.002236649626865983, + 0.005994497332721949, + 0.00890191737562418, + 0.004426663275808096, + 0.005608671810477972, + 0.006242021918296814, + 0.005490134470164776, + 0.0029713204130530357, + 0.0032874944154173136, + 0.0044142622500658035, + 0.007020223420113325, + 0.0046464428305625916, + 0.004143571946769953, + 0.00563229713588953, + 0.0031711359042674303, + 0.007073892746120691, + 0.010936672799289227, + 0.013057555072009563, + 0.0092606907710433, + 0.010798202827572823, + 0.02324104495346546, + 0.010372894816100597, + 0.010311954654753208, + 0.0046888780780136585, + 0.0027596920263022184, + 0.004420545883476734, + 0.009103582240641117, + 0.005251085851341486, + 0.005647944286465645, + 0.010976252146065235, + 0.0038181317504495382, + 0.004605874419212341, + 0.0023628142662346363, + 0.006429153960198164, + 0.00418391078710556, + 0.004178464878350496, + 0.00594996428117156, + 0.0028954101726412773, + 0.006146301049739122, + 0.036763645708560944, + 0.021419866010546684, + 0.06720233708620071, + 0.04391727223992348, + 0.04737154021859169, + 0.0684279352426529, + 0.05561714619398117, + 0.06675998866558075, + 0.005974511615931988, + 0.00619383854791522, + 0.010071889497339725, + 0.008844382129609585, + 0.006407466251403093, + 0.010425489395856857, + 0.008835348300635815, + 0.007388204801827669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.013282505795359612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641134783625603 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641134783625603, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641134783625603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17106685042381287, + 0.08127021044492722, + 0.10281840711832047, + 0.04365606606006622, + 0.050806254148483276, + 0.0756315365433693, + 0.049166250973939896, + 0.16585253179073334, + 0.21676360070705414, + 0.04296831786632538 + ], + "credit_map": [ + 0.016034971922636032, + 0.01285781804472208, + 0.02825433574616909, + 0.03236008808016777, + 0.021848510950803757, + 0.02401890605688095, + 0.015965618193149567, + 0.01972658932209015, + 0.010891308076679707, + 0.006595275364816189, + 0.016043195500969887, + 0.007563670631498098, + 0.010510370135307312, + 0.012336501851677895, + 0.008132525719702244, + 0.009197362698614597, + 0.010177375748753548, + 0.004940714221447706, + 0.01770324632525444, + 0.011263753287494183, + 0.010519278235733509, + 0.019091207534074783, + 0.013834916986525059, + 0.015287901274859905, + 0.005649683531373739, + 0.0021979468874633312, + 0.00728262634947896, + 0.005546519532799721, + 0.005994460079818964, + 0.006395550910383463, + 0.00522844772785902, + 0.005360827781260014, + 0.004221782088279724, + 0.004023425746709108, + 0.00561390770599246, + 0.011487285606563091, + 0.004572669044137001, + 0.006463277619332075, + 0.009867673739790916, + 0.004556222818791866, + 0.005502799525856972, + 0.005821734666824341, + 0.008665964938700199, + 0.005469677969813347, + 0.015545309521257877, + 0.012621221132576466, + 0.01508873887360096, + 0.00691608339548111, + 0.004811860155314207, + 0.0032399084884673357, + 0.006032107397913933, + 0.009280500002205372, + 0.00579189183190465, + 0.006025938782840967, + 0.008787239901721478, + 0.0051967971958220005, + 0.01968594267964363, + 0.010915474034845829, + 0.02578677237033844, + 0.015511738136410713, + 0.019875433295965195, + 0.03358890488743782, + 0.008554324507713318, + 0.03193393722176552, + 0.020554721355438232, + 0.012117450125515461, + 0.03580908477306366, + 0.024200180545449257, + 0.028558721765875816, + 0.03561025112867355, + 0.02667076140642166, + 0.03324240446090698, + 0.0036031848285347223, + 0.002936367643997073, + 0.005368941463530064, + 0.00566439051181078, + 0.005986830685287714, + 0.006674847565591335, + 0.0072099133394658566, + 0.005523836240172386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.007641134783625603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033608973026276 + ], + "opposite_logprob_tokens": [ + -2.3503360748291016 + ], + "logprob_reward": -0.10033608973026276, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033608973026276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12604252994060516, + 0.10827852040529251, + 0.1521768420934677, + 0.084499791264534, + 0.04490083456039429, + 0.20515413582324982, + 0.07457129657268524, + 0.04073010012507439, + 0.11159098893404007, + 0.05205490067601204 + ], + "credit_map": [ + 0.011746399104595184, + 0.009481816552579403, + 0.012921720743179321, + 0.020794671028852463, + 0.016652468591928482, + 0.021721387282013893, + 0.01729232631623745, + 0.015431755222380161, + 0.009434320032596588, + 0.007261908147484064, + 0.017248591408133507, + 0.010874430648982525, + 0.01735260710120201, + 0.01808210276067257, + 0.015341506339609623, + 0.012683062814176083, + 0.022442461922764778, + 0.010279438458383083, + 0.02477364055812359, + 0.02240918017923832, + 0.014093817211687565, + 0.02006904035806656, + 0.016264982521533966, + 0.021844301372766495, + 0.008252699859440327, + 0.005780396051704884, + 0.007864028215408325, + 0.01852726750075817, + 0.011155986227095127, + 0.010441756807267666, + 0.014250112697482109, + 0.008227554149925709, + 0.003563651582226157, + 0.0027585451025515795, + 0.004108813591301441, + 0.006528737023472786, + 0.006005949340760708, + 0.009108994156122208, + 0.007957346737384796, + 0.004868800286203623, + 0.015358000993728638, + 0.013375705108046532, + 0.019511545076966286, + 0.01464627031236887, + 0.041266169399023056, + 0.029666606336832047, + 0.05313405022025108, + 0.01819581724703312, + 0.006149498280137777, + 0.004552002996206284, + 0.00884590856730938, + 0.014381593093276024, + 0.00912535935640335, + 0.008809727616608143, + 0.015704281628131866, + 0.007002938538789749, + 0.0053815594874322414, + 0.002869712421670556, + 0.005555577110499144, + 0.006545924115926027, + 0.0051727998070418835, + 0.00555048743262887, + 0.005026535596698523, + 0.00462750531733036, + 0.010243120603263378, + 0.006201846990734339, + 0.018232950940728188, + 0.013093804009258747, + 0.013388806022703648, + 0.017408009618520737, + 0.015644259750843048, + 0.017378194257616997, + 0.004808122757822275, + 0.00488371541723609, + 0.005713004618883133, + 0.007089424878358841, + 0.0074567669071257114, + 0.006944629829376936, + 0.010450931265950203, + 0.00470831198617816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.10033608973026276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418777834624052 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.01418777834624052, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418777834624052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1530095785856247, + 0.07876941561698914, + 0.11478982120752335, + 0.09060300141572952, + 0.090880386531353, + 0.11575488746166229, + 0.10457352548837662, + 0.11683250218629837, + 0.09461929649114609, + 0.04016760364174843 + ], + "credit_map": [ + 0.011803258210420609, + 0.010147299617528915, + 0.020245211198925972, + 0.028241561725735664, + 0.01617615856230259, + 0.02767784334719181, + 0.021948283538222313, + 0.016769971698522568, + 0.009459900669753551, + 0.006408455781638622, + 0.012934327125549316, + 0.010345890186727047, + 0.010609263554215431, + 0.011712230741977692, + 0.009007866494357586, + 0.00829147920012474, + 0.01053300779312849, + 0.00649101659655571, + 0.016932586207985878, + 0.013357534073293209, + 0.011863531544804573, + 0.02202230878174305, + 0.013495459221303463, + 0.0200943760573864, + 0.008054183796048164, + 0.005619130562990904, + 0.010243339464068413, + 0.021002234891057014, + 0.008117815479636192, + 0.011426372453570366, + 0.016514405608177185, + 0.009625516831874847, + 0.006915474310517311, + 0.007285146974027157, + 0.009761154651641846, + 0.019238630309700966, + 0.012880347669124603, + 0.012946094386279583, + 0.01474832370877266, + 0.007105210795998573, + 0.008197503164410591, + 0.007252844516187906, + 0.010988582856953144, + 0.010575060732662678, + 0.02647254429757595, + 0.01761140301823616, + 0.024452345445752144, + 0.010204596444964409, + 0.008267723955214024, + 0.005896190647035837, + 0.011289079673588276, + 0.019487299025058746, + 0.012340942397713661, + 0.012826211750507355, + 0.025691665709018707, + 0.00877441093325615, + 0.018149549141526222, + 0.006733378861099482, + 0.01816028542816639, + 0.015342921949923038, + 0.01182693149894476, + 0.019365476444363594, + 0.009003836661577225, + 0.018250130116939545, + 0.007542865350842476, + 0.005225055851042271, + 0.013972203247249126, + 0.012671023607254028, + 0.0103093096986413, + 0.01580561138689518, + 0.014704071916639805, + 0.014389156363904476, + 0.003281370969489217, + 0.0035061584785580635, + 0.005127663724124432, + 0.0038579613901674747, + 0.00557145569473505, + 0.010240931063890457, + 0.004784454591572285, + 0.0037976084277033806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.01418777834624052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978617139160633 + ], + "opposite_logprob_tokens": [ + -3.5297861099243164 + ], + "logprob_reward": -0.02978617139160633, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978617139160633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22290606796741486, + 0.0675867572426796, + 0.23483730852603912, + 0.06001291424036026, + 0.1170775294303894, + 0.06870624423027039, + 0.08096423745155334, + 0.0534881092607975, + 0.03567542880773544, + 0.058745406568050385 + ], + "credit_map": [ + 0.022153262048959732, + 0.016271857544779778, + 0.03191523626446724, + 0.03611965477466583, + 0.02590092085301876, + 0.031252261251211166, + 0.030527306720614433, + 0.028765559196472168, + 0.01065001543611288, + 0.006024930626153946, + 0.009635224007070065, + 0.0073694935999810696, + 0.0077292160131037235, + 0.010961313731968403, + 0.006583919283002615, + 0.008632647804915905, + 0.025426026433706284, + 0.012919336557388306, + 0.02801169641315937, + 0.03599033132195473, + 0.023651227355003357, + 0.03781367465853691, + 0.03763607516884804, + 0.03338893502950668, + 0.005954104010015726, + 0.0040817237459123135, + 0.009832553565502167, + 0.00732621643692255, + 0.00799296423792839, + 0.009279344230890274, + 0.007553615607321262, + 0.007992392405867577, + 0.011992435902357101, + 0.0083066551014781, + 0.017821386456489563, + 0.013354158960282803, + 0.018498526886105537, + 0.01629607193171978, + 0.015448136255145073, + 0.015360159799456596, + 0.004986039362847805, + 0.005215499084442854, + 0.008415955118834972, + 0.007552207913249731, + 0.013226928189396858, + 0.010284597985446453, + 0.012324556708335876, + 0.006700461730360985, + 0.007397411856800318, + 0.0052341921254992485, + 0.008259263820946217, + 0.015251483768224716, + 0.009641802869737148, + 0.010028810240328312, + 0.016811246052384377, + 0.008340028114616871, + 0.007180340122431517, + 0.003095183288678527, + 0.009965427219867706, + 0.0077804457396268845, + 0.0057172323577106, + 0.006920202635228634, + 0.0044203163124620914, + 0.008408959954977036, + 0.0033139323350042105, + 0.0020869020372629166, + 0.004337781574577093, + 0.005174500867724419, + 0.004632401745766401, + 0.005361148621886969, + 0.006044770125299692, + 0.004723990801721811, + 0.006276905070990324, + 0.00511274253949523, + 0.007260062266141176, + 0.007794321049004793, + 0.007674342021346092, + 0.010291999205946922, + 0.0077049341052770615, + 0.006630096584558487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02978617139160633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crisp, static, top-down, high-angle shot offers a perfect bird's-eye view of a vibrantly bouncing ball as it rebounds powerfully across a clean, expansive wooden tabletop. The table, crafted from warm, honey-toned oak, polished to a soft satin, reveals its delicate, linear grain under the bright, even, and diffuse overhead lighting. The ball itself is a chaotic swirl of electric blue, neon yellow, and hot pink, fused together into a glossy, high-density polymer captured in sharp, circular highlights. With each rhythmic impact, the ball compresses almost imperceptibly against the hardwood before rebounding into the air, its rapid rotation blurring the colors into dynamic streaks as it traces a series of graceful, diminishing arcs across the frame. A soft, perfectly circular shadow dances directly below; its dark twin contracts to a point upon impact and expands as the ball soars, creating a hypnotic, playful atmosphere—a pure and minimalist study in motion, color, and physics.", + "seed": 31999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_3649_seed_31999_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.040424348786473274, + "vlm_logprob": -0.040424348786473274, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.040424348786473274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10277016460895538 + ], + "opposite_logprob_tokens": [ + -2.4777700901031494 + ], + "logprob_reward": -0.10277016460895538, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10277016460895538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1722298562526703, + 0.08704731613397598, + 0.09696925431489944, + 0.0706750825047493, + 0.10257623344659805, + 0.04057803377509117, + 0.19807446002960205, + 0.06964710354804993, + 0.06819789856672287, + 0.09400471299886703 + ], + "credit_map": [ + 0.016830895096063614, + 0.019185269251465797, + 0.026771370321512222, + 0.02072742208838463, + 0.016573773697018623, + 0.018579360097646713, + 0.014038230292499065, + 0.03952356427907944, + 0.0072963666170835495, + 0.009708120487630367, + 0.013170371763408184, + 0.009678201749920845, + 0.006102877203375101, + 0.01087389700114727, + 0.007625206373631954, + 0.022592274472117424, + 0.011161412112414837, + 0.010321507230401039, + 0.020337330177426338, + 0.01799219846725464, + 0.008548296056687832, + 0.007242721505463123, + 0.012336958199739456, + 0.009028840810060501, + 0.007433574181050062, + 0.007257705554366112, + 0.01031874492764473, + 0.010849458165466785, + 0.00801131222397089, + 0.006549284793436527, + 0.008478970266878605, + 0.011776039376854897, + 0.0107410978525877, + 0.011844816617667675, + 0.024064188823103905, + 0.011781110428273678, + 0.008796749636530876, + 0.00827683787792921, + 0.010702448897063732, + 0.016368988901376724, + 0.003923414275050163, + 0.002181774005293846, + 0.008198305033147335, + 0.0044952635653316975, + 0.003918653354048729, + 0.0028110311832278967, + 0.002890141448006034, + 0.012159451842308044, + 0.015173078514635563, + 0.0051630353555083275, + 0.07330753654241562, + 0.014153223484754562, + 0.007433857303112745, + 0.011488299816846848, + 0.011227170005440712, + 0.060128286480903625, + 0.010822015814483166, + 0.010888430289924145, + 0.01094143372029066, + 0.006947427522391081, + 0.007869338616728783, + 0.007183202542364597, + 0.007133258040994406, + 0.007862001657485962, + 0.00663650780916214, + 0.009155664592981339, + 0.007799621671438217, + 0.01051346119493246, + 0.012033754959702492, + 0.004599707666784525, + 0.00992052722722292, + 0.007538657169789076, + 0.012476028874516487, + 0.009916802868247032, + 0.02179182693362236, + 0.006794172339141369, + 0.007558668963611126, + 0.006845777854323387, + 0.0052643115632236, + 0.023357126861810684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.10277016460895538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721499145030975 + ], + "opposite_logprob_tokens": [ + -3.408721446990967 + ], + "logprob_reward": -0.033721499145030975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721499145030975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13519123196601868, + 0.05791756883263588, + 0.07243096083402634, + 0.11206331104040146, + 0.08034489303827286, + 0.05881607532501221, + 0.08432933688163757, + 0.24003763496875763, + 0.0342865027487278, + 0.12458249181509018 + ], + "credit_map": [ + 0.013217661529779434, + 0.02596919611096382, + 0.01561628095805645, + 0.01938593201339245, + 0.014771399088203907, + 0.013161448761820793, + 0.015951460227370262, + 0.017117854207754135, + 0.005758399609476328, + 0.009207780472934246, + 0.009748908691108227, + 0.006533205043524504, + 0.004304223228245974, + 0.009142013266682625, + 0.005258444231003523, + 0.007964590564370155, + 0.007346950005739927, + 0.007359596900641918, + 0.01458836067467928, + 0.00921429879963398, + 0.007410756777971983, + 0.007441670633852482, + 0.012387547641992569, + 0.006681776139885187, + 0.013490743935108185, + 0.01283005066215992, + 0.02220749482512474, + 0.019327159970998764, + 0.007937828078866005, + 0.018890924751758575, + 0.005524598993360996, + 0.011854509823024273, + 0.013740817084908485, + 0.00842061173170805, + 0.011466015130281448, + 0.013042176142334938, + 0.008926083333790302, + 0.008726990781724453, + 0.0090707428753376, + 0.006951451767235994, + 0.004751425236463547, + 0.003807268338277936, + 0.013186177238821983, + 0.006534548941999674, + 0.00498121976852417, + 0.003280547447502613, + 0.005114533007144928, + 0.01716035045683384, + 0.006196248345077038, + 0.0032706770580261946, + 0.027656521648168564, + 0.007628344930708408, + 0.0051711625419557095, + 0.0054652211256325245, + 0.004588518291711807, + 0.02435263991355896, + 0.025749417021870613, + 0.024130674079060555, + 0.06343325972557068, + 0.017188452184200287, + 0.018295777961611748, + 0.02059665508568287, + 0.01063713151961565, + 0.06000624597072601, + 0.0031840854790061712, + 0.006653327029198408, + 0.004779967479407787, + 0.0041556451469659805, + 0.004907375201582909, + 0.0019193835323676467, + 0.0040332297794520855, + 0.004653488751500845, + 0.010455960407853127, + 0.007264041807502508, + 0.04245354235172272, + 0.005809079390019178, + 0.005182900000363588, + 0.005545163527131081, + 0.00409551989287138, + 0.04377627745270729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.033721499145030975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305077105760574 + ], + "opposite_logprob_tokens": [ + -3.7733051776885986 + ], + "logprob_reward": -0.023305077105760574, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305077105760574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0808316320180893, + 0.09206443279981613, + 0.16451242566108704, + 0.06144626811146736, + 0.08255108445882797, + 0.029830720275640488, + 0.12665164470672607, + 0.11827161908149719, + 0.08763916790485382, + 0.15620099008083344 + ], + "credit_map": [ + 0.008872395381331444, + 0.008668821305036545, + 0.00968516618013382, + 0.0111560532823205, + 0.008491055108606815, + 0.0072073196060955524, + 0.013129759579896927, + 0.013621059246361256, + 0.00804400909692049, + 0.010443289764225483, + 0.012407279573380947, + 0.00964276771992445, + 0.004183325916528702, + 0.010333128273487091, + 0.003480559680610895, + 0.03353006765246391, + 0.012633624486625195, + 0.02244296669960022, + 0.04379599913954735, + 0.019977249205112457, + 0.024720272049307823, + 0.012113088741898537, + 0.016791673377156258, + 0.012037565000355244, + 0.006485166493803263, + 0.007986013777554035, + 0.010134117677807808, + 0.008226139470934868, + 0.009449316188693047, + 0.006324595306068659, + 0.005302070174366236, + 0.0075388564728200436, + 0.007520826067775488, + 0.009279769845306873, + 0.019549090415239334, + 0.008955958299338818, + 0.010915946215391159, + 0.00702426303178072, + 0.006860201712697744, + 0.012445027939975262, + 0.0028009950183331966, + 0.002362314146012068, + 0.003981498070061207, + 0.004138292744755745, + 0.004630999639630318, + 0.002970987930893898, + 0.00389356492087245, + 0.005052069202065468, + 0.00913348887115717, + 0.003779452294111252, + 0.04450060427188873, + 0.008927187882363796, + 0.0062683275900781155, + 0.006899740546941757, + 0.007673818152397871, + 0.03946901485323906, + 0.009455012157559395, + 0.006646488793194294, + 0.04144957289099693, + 0.005585309583693743, + 0.0035959878005087376, + 0.008819438517093658, + 0.0031711754854768515, + 0.03954864665865898, + 0.014161846600472927, + 0.010412358678877354, + 0.010282212868332863, + 0.00697116507217288, + 0.01581289991736412, + 0.00859767384827137, + 0.00855623185634613, + 0.012844778597354889, + 0.01720096729695797, + 0.006695773918181658, + 0.05238780751824379, + 0.007574622984975576, + 0.0057134488597512245, + 0.007274678908288479, + 0.006883162539452314, + 0.05247052386403084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.023305077105760574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14571474492549896, + 0.2094932496547699, + 0.1441204994916916, + 0.08204765617847443, + 0.0666976049542427, + 0.0646999329328537, + 0.06249115616083145, + 0.046377744525671005, + 0.11662802845239639, + 0.06172945722937584 + ], + "credit_map": [ + 0.01572389155626297, + 0.024518776684999466, + 0.017762653529644012, + 0.018056390807032585, + 0.022681469097733498, + 0.013354968279600143, + 0.01590762846171856, + 0.01770895905792713, + 0.016787448897957802, + 0.023064501583576202, + 0.02973303571343422, + 0.022894714027643204, + 0.009058348834514618, + 0.026441674679517746, + 0.008171599358320236, + 0.07334192842245102, + 0.018978537991642952, + 0.01400228776037693, + 0.02221480943262577, + 0.024480069056153297, + 0.021405570209026337, + 0.010845853947103024, + 0.016471732407808304, + 0.015721643343567848, + 0.009784075431525707, + 0.010594579391181469, + 0.013275439850986004, + 0.013551882468163967, + 0.0068546803668141365, + 0.014176584780216217, + 0.0054409801959991455, + 0.00836943369358778, + 0.004617925267666578, + 0.007731310091912746, + 0.01838643290102482, + 0.007575675379484892, + 0.006876280531287193, + 0.006256886757910252, + 0.006199694704264402, + 0.009053398855030537, + 0.007075258996337652, + 0.0034515894949436188, + 0.008989615365862846, + 0.00816644262522459, + 0.006790008395910263, + 0.005484033841639757, + 0.005274862051010132, + 0.019468117505311966, + 0.00530193792656064, + 0.006470549385994673, + 0.015411081723868847, + 0.005992501974105835, + 0.006794526241719723, + 0.004088898189365864, + 0.005588881205767393, + 0.012842780910432339, + 0.0050511728040874004, + 0.008605598472058773, + 0.00700021255761385, + 0.004432490095496178, + 0.008249261416494846, + 0.003676967229694128, + 0.003617005655542016, + 0.00574503606185317, + 0.013252577744424343, + 0.018564043566584587, + 0.013206327334046364, + 0.01569175161421299, + 0.015535098500549793, + 0.008262953720986843, + 0.020427504554390907, + 0.01168775837868452, + 0.007792128715664148, + 0.010511943139135838, + 0.009315013885498047, + 0.00723186694085598, + 0.007521725259721279, + 0.004803426098078489, + 0.008072095923125744, + 0.006481258664280176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.012535706162452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978929691016674 + ], + "opposite_logprob_tokens": [ + -3.5297892093658447 + ], + "logprob_reward": -0.02978929691016674, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978929691016674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1830572485923767, + 0.10012967884540558, + 0.3496299684047699, + 0.0399499274790287, + 0.04580868035554886, + 0.04339354485273361, + 0.02525903843343258, + 0.055257707834243774, + 0.08979538083076477, + 0.06771887093782425 + ], + "credit_map": [ + 0.017180107533931732, + 0.03353298082947731, + 0.024808716028928757, + 0.022901330143213272, + 0.023708706721663475, + 0.02103973738849163, + 0.014432098716497421, + 0.025453560054302216, + 0.007702352479100227, + 0.010972938500344753, + 0.019918041303753853, + 0.00991377979516983, + 0.006151870358735323, + 0.012103988789021969, + 0.0064719123765826225, + 0.026894791051745415, + 0.03158652037382126, + 0.041996926069259644, + 0.06396228820085526, + 0.06200472265481949, + 0.03486097231507301, + 0.03999362140893936, + 0.044902361929416656, + 0.030322538688778877, + 0.0039242287166416645, + 0.004771909676492214, + 0.009452970698475838, + 0.006553828716278076, + 0.002283821115270257, + 0.006323608104139566, + 0.0026287087239325047, + 0.0040108500979840755, + 0.006180468015372753, + 0.004928403999656439, + 0.008695187047123909, + 0.006633722223341465, + 0.003961157985031605, + 0.005257826764136553, + 0.0044602504931390285, + 0.005691660568118095, + 0.004754136316478252, + 0.002573251025751233, + 0.007255710661411285, + 0.004850116092711687, + 0.003506803186610341, + 0.004242542665451765, + 0.003059905953705311, + 0.013151074759662151, + 0.0028921037446707487, + 0.0024874266237020493, + 0.004356231074780226, + 0.003452718025073409, + 0.0030659246258437634, + 0.0024796926882117987, + 0.002019386738538742, + 0.004505551420152187, + 0.006085018161684275, + 0.0031414837576448917, + 0.01829097792506218, + 0.0031099047046154737, + 0.002536516170948744, + 0.003475398989394307, + 0.0028425343334674835, + 0.01577587239444256, + 0.008025536313652992, + 0.01355494000017643, + 0.01090649701654911, + 0.01236466784030199, + 0.017418840900063515, + 0.006038645748049021, + 0.00998070277273655, + 0.011505542322993279, + 0.008817709982395172, + 0.011065831407904625, + 0.010396786034107208, + 0.006063627079129219, + 0.007559618446975946, + 0.007586559280753136, + 0.005815009120851755, + 0.010413718409836292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02978929691016674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "crisp, static, top-down, high-angle shot offers a perfect bird's-eye view of a vibrantly bouncing ball as it rebounds powerfully across a clean, expansive wooden tabletop. The table, crafted from warm, honey-toned oak, polished to a soft satin, reveals its delicate, linear grain under the bright, even, and diffuse overhead lighting. The ball itself is a chaotic swirl of electric blue, neon yellow, and hot pink, fused together into a glossy, high-density polymer captured in sharp, circular highlights. With each rhythmic impact, the ball compresses almost imperceptibly against the hardwood before rebounding into the air, its rapid rotation blurring the colors into dynamic streaks as it traces a series of graceful, diminishing arcs across the frame. A soft, perfectly circular shadow dances directly below; its dark twin contracts to a point upon impact and expands as the ball soars, creating a hypnotic, playful atmosphere—a pure and minimalist study in motion, color, and physics.", + "seed": 32000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_4/index_3649_seed_32000_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028833528980612755, + "vlm_logprob": -0.028833528980612755, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.028833528980612755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05059479549527168 + ], + "opposite_logprob_tokens": [ + -3.1755948066711426 + ], + "logprob_reward": -0.05059479549527168, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05059479549527168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08437979966402054, + 0.06488236784934998, + 0.02780533954501152, + 0.07588719576597214, + 0.05668229982256889, + 0.1425650417804718, + 0.2814342677593231, + 0.21399284899234772, + 0.023528030142188072, + 0.028842858970165253 + ], + "credit_map": [ + 0.006107841618359089, + 0.0052365451119840145, + 0.027330176904797554, + 0.006309696938842535, + 0.004938963800668716, + 0.008743158541619778, + 0.006794532295316458, + 0.01891889050602913, + 0.0061545604839921, + 0.004779184237122536, + 0.011025164276361465, + 0.004845400340855122, + 0.00708157941699028, + 0.010681049898266792, + 0.004451765678822994, + 0.01586366817355156, + 0.0026848616544157267, + 0.0024826100561767817, + 0.006807493511587381, + 0.0025878536980599165, + 0.002950386144220829, + 0.0023215864785015583, + 0.0021869835909456015, + 0.005783564411103725, + 0.005860020872205496, + 0.0058839754201471806, + 0.024972792714834213, + 0.004310812801122665, + 0.0037924768403172493, + 0.00608469732105732, + 0.004838740918785334, + 0.02014368586242199, + 0.004091140814125538, + 0.003943371120840311, + 0.01810264214873314, + 0.0033327967394143343, + 0.004262917675077915, + 0.0044821202754974365, + 0.0032712023239582777, + 0.015196109190583229, + 0.008162345737218857, + 0.008352285251021385, + 0.047762610018253326, + 0.0070731849409639835, + 0.006909537594765425, + 0.011265050619840622, + 0.007146551739424467, + 0.04589347168803215, + 0.017339833080768585, + 0.021344969049096107, + 0.1137097179889679, + 0.015039012767374516, + 0.018422652035951614, + 0.014726420864462852, + 0.011111906729638577, + 0.06973977386951447, + 0.012958814390003681, + 0.016843954101204872, + 0.06510820239782333, + 0.015079540200531483, + 0.010247881524264812, + 0.02485310286283493, + 0.011075473390519619, + 0.05782587453722954, + 0.0028065547812730074, + 0.0025112370494753122, + 0.0031524640507996082, + 0.0029152368661016226, + 0.0024080148432403803, + 0.0035304257180541754, + 0.0022620651870965958, + 0.003942030947655439, + 0.0024707072880119085, + 0.002920052269473672, + 0.0055985101498663425, + 0.0027111335657536983, + 0.0025255782529711723, + 0.0043784151785075665, + 0.0015601367922499776, + 0.006678326521068811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.05059479549527168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201411046087742 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201411046087742, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201411046087742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23179098963737488, + 0.17323124408721924, + 0.08873723447322845, + 0.15487267076969147, + 0.15816046297550201, + 0.028963571414351463, + 0.06259406358003616, + 0.03354092314839363, + 0.006767801474779844, + 0.061341024935245514 + ], + "credit_map": [ + 0.01496740523725748, + 0.01244038064032793, + 0.08476084470748901, + 0.010891149751842022, + 0.011141720227897167, + 0.020457597449421883, + 0.009971598163247108, + 0.06716030091047287, + 0.023779405280947685, + 0.023267222568392754, + 0.02234623208642006, + 0.01321425847709179, + 0.01672350987792015, + 0.026142604649066925, + 0.01535580214112997, + 0.03240222483873367, + 0.005993370898067951, + 0.00628997664898634, + 0.030271464958786964, + 0.00498672341927886, + 0.006415477488189936, + 0.0067592766135931015, + 0.005009955260902643, + 0.023010995239019394, + 0.01004953496158123, + 0.005484286230057478, + 0.06381535530090332, + 0.00603965250775218, + 0.00617254851385951, + 0.0083078658208251, + 0.0041554891504347324, + 0.050847962498664856, + 0.008601028472185135, + 0.010203294456005096, + 0.062076468020677567, + 0.0050519309006631374, + 0.007736142259091139, + 0.00871626753360033, + 0.0038973367772996426, + 0.051878008991479874, + 0.0025510420091450214, + 0.0014370543649420142, + 0.009905575774610043, + 0.0016167473513633013, + 0.0014556492678821087, + 0.0027373959310352802, + 0.0015985502395778894, + 0.007661559619009495, + 0.0037335322704166174, + 0.005510392598807812, + 0.025621948763728142, + 0.0025135776959359646, + 0.0031388946808874607, + 0.002865041373297572, + 0.0018841263372451067, + 0.017326558008790016, + 0.0021334588527679443, + 0.0027470597997307777, + 0.009912492707371712, + 0.001887979218736291, + 0.0018562021432444453, + 0.004101370461285114, + 0.0019028810784220695, + 0.008999479934573174, + 0.0007435825536958873, + 0.0006219211500138044, + 0.0011821691878139973, + 0.0008794798632152379, + 0.0006037498824298382, + 0.0009332955232821405, + 0.0005008850712329149, + 0.0013027183013036847, + 0.0048221019096672535, + 0.004255224019289017, + 0.017077244818210602, + 0.0027358236256986856, + 0.003600846277549863, + 0.00792243704199791, + 0.003075368469581008, + 0.01785198226571083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014201411046087742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662614077329636 + ], + "opposite_logprob_tokens": [ + -3.0486626625061035 + ], + "logprob_reward": -0.048662614077329636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662614077329636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2468574196100235, + 0.21432656049728394, + 0.08522346615791321, + 0.06115604564547539, + 0.1091575175523758, + 0.022623157128691673, + 0.0432218499481678, + 0.05019279941916466, + 0.007627103943377733, + 0.15961408615112305 + ], + "credit_map": [ + 0.015616336837410927, + 0.012785828672349453, + 0.0932299792766571, + 0.009832989424467087, + 0.00968056358397007, + 0.022424474358558655, + 0.007377741392701864, + 0.07590949535369873, + 0.030953537672758102, + 0.02961273118853569, + 0.02792595699429512, + 0.016249312087893486, + 0.020960217341780663, + 0.03304193541407585, + 0.016049468889832497, + 0.039533380419015884, + 0.005803184118121862, + 0.004718915559351444, + 0.025163520127534866, + 0.00898593571037054, + 0.007244091015309095, + 0.005659473594278097, + 0.006220248062163591, + 0.021428091451525688, + 0.00732550211250782, + 0.010523313656449318, + 0.010628372430801392, + 0.005422133021056652, + 0.007478434592485428, + 0.008173017762601376, + 0.004406503867357969, + 0.007198763079941273, + 0.0056918589398264885, + 0.006280779838562012, + 0.04001087322831154, + 0.0053517185151577, + 0.006520641036331654, + 0.0058680507354438305, + 0.0053542610257864, + 0.03407933562994003, + 0.0016175630735233426, + 0.002510202582925558, + 0.0045923758298158646, + 0.0019728229381144047, + 0.00187463138718158, + 0.002694013761356473, + 0.002213865052908659, + 0.005147682502865791, + 0.002811077982187271, + 0.003106364281848073, + 0.01809791661798954, + 0.0022192620672285557, + 0.0031472898554056883, + 0.0023546128068119287, + 0.0014918739907443523, + 0.009993448853492737, + 0.002972291549667716, + 0.0036032970529049635, + 0.016798192635178566, + 0.0031485557556152344, + 0.001909301383420825, + 0.0051807439886033535, + 0.0022848169319331646, + 0.014295597560703754, + 0.0006665397086180747, + 0.00069514597998932, + 0.001294159097597003, + 0.0013227418530732393, + 0.0006725491839461029, + 0.000897770281881094, + 0.0008039119420573115, + 0.0012742856051772833, + 0.014039498753845692, + 0.013940279372036457, + 0.0389421246945858, + 0.009145972318947315, + 0.009793281555175781, + 0.02262224815785885, + 0.007093360647559166, + 0.04403730109333992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.048662614077329636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531938962638378 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531938962638378, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531938962638378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.313154935836792, + 0.1657293289899826, + 0.033504657447338104, + 0.04503403231501579, + 0.14098000526428223, + 0.0760805606842041, + 0.020742256194353104, + 0.08667653799057007, + 0.07330338656902313, + 0.04479428753256798 + ], + "credit_map": [ + 0.018117204308509827, + 0.01694803684949875, + 0.1116417869925499, + 0.016931746155023575, + 0.01404702290892601, + 0.03003651648759842, + 0.014936210587620735, + 0.09049642086029053, + 0.020566057413816452, + 0.020853284746408463, + 0.02250814624130726, + 0.01221507415175438, + 0.02072552591562271, + 0.021564044058322906, + 0.016408732160925865, + 0.03088846616446972, + 0.002939315279945731, + 0.00331212324090302, + 0.007877065800130367, + 0.003714577993378043, + 0.0036992018576711416, + 0.003002818441018462, + 0.0036191625986248255, + 0.005340390373021364, + 0.005610042251646519, + 0.004538035020232201, + 0.010879662819206715, + 0.0034296743106096983, + 0.002985498635098338, + 0.006785391829907894, + 0.004369029309600592, + 0.006436699070036411, + 0.009527773596346378, + 0.008686025626957417, + 0.044876884669065475, + 0.00917530246078968, + 0.010407650843262672, + 0.012686706148087978, + 0.00876747164875269, + 0.03685219958424568, + 0.0052171070128679276, + 0.0037923529744148254, + 0.024974973872303963, + 0.004585877526551485, + 0.003438608953729272, + 0.0059919157065451145, + 0.00485166534781456, + 0.023228056728839874, + 0.002385005820542574, + 0.0031570501159876585, + 0.0031142979860305786, + 0.0024290287401527166, + 0.002552577294409275, + 0.002773209474980831, + 0.0021083939354866743, + 0.0022226914297789335, + 0.007527925074100494, + 0.007297984790056944, + 0.01978103071451187, + 0.011908629909157753, + 0.006147222127765417, + 0.009214133024215698, + 0.007350423373281956, + 0.017449188977479935, + 0.0067421263083815575, + 0.007242102641612291, + 0.018513552844524384, + 0.009588993154466152, + 0.006337474565953016, + 0.006739435717463493, + 0.008308134973049164, + 0.009831566363573074, + 0.0031805643811821938, + 0.004317845217883587, + 0.00926564447581768, + 0.004706702195107937, + 0.00418127654120326, + 0.0060370867140591145, + 0.002530205762013793, + 0.010574962012469769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012531938962638378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static, high-angle top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface a clean, expansive wooden tabletop made of warm, honey-toned oak with visible linear grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ball a glossy, high-density polymer with a chaotic swirl of electric blue, neon yellow, and hot pink colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball bounce rhythmically across the table, compressing upon impact and rebounding into a series of diminishing arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, perfectly circular shadow directly below the ball that contracts to a point on impact and expands as the ball rises?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25640353560447693, + 0.0647733062505722, + 0.03705304116010666, + 0.1029028668999672, + 0.021945863962173462, + 0.3425334095954895, + 0.038889020681381226, + 0.06213666871190071, + 0.02292276918888092, + 0.0504394955933094 + ], + "credit_map": [ + 0.01748119853436947, + 0.012932650744915009, + 0.094721220433712, + 0.012648070231080055, + 0.011498703621327877, + 0.018726013600826263, + 0.015486110001802444, + 0.07290956377983093, + 0.006219266448169947, + 0.006192986387759447, + 0.011767196469008923, + 0.0053474134765565395, + 0.006090882699936628, + 0.007423365488648415, + 0.007151634432375431, + 0.014580555260181427, + 0.0031605090480297804, + 0.0023607867769896984, + 0.012852732092142105, + 0.0021261260844767094, + 0.0025145127438008785, + 0.0033334495965391397, + 0.002145071979612112, + 0.008559847250580788, + 0.006457002367824316, + 0.004938998259603977, + 0.038557227700948715, + 0.004336575046181679, + 0.00599460257217288, + 0.006111687980592251, + 0.00564019987359643, + 0.03086656890809536, + 0.002141448901966214, + 0.0031106199603527784, + 0.0032566958107054234, + 0.002510558580979705, + 0.0031298340763896704, + 0.002395838964730501, + 0.0028943924698978662, + 0.0025064728688448668, + 0.018588779494166374, + 0.010065504349768162, + 0.14533932507038116, + 0.011761974543333054, + 0.007504974491894245, + 0.017934588715434074, + 0.00787433236837387, + 0.12346389144659042, + 0.00275003002025187, + 0.0034622810781002045, + 0.013160430826246738, + 0.0022532630246132612, + 0.0020428758580237627, + 0.003658983390778303, + 0.0017507264856249094, + 0.009810429997742176, + 0.003720999462530017, + 0.005333272274583578, + 0.018157007172703743, + 0.004563771653920412, + 0.0030071469955146313, + 0.007176785729825497, + 0.003772903233766556, + 0.016404779627919197, + 0.0027544170152395964, + 0.002854844555258751, + 0.004310802556574345, + 0.00255505065433681, + 0.0023685398045927286, + 0.0029547461308538914, + 0.002517202403396368, + 0.002607164438813925, + 0.004731173161417246, + 0.004854774102568626, + 0.009449278935790062, + 0.0044005680829286575, + 0.003944795113056898, + 0.007504334673285484, + 0.003508353140205145, + 0.012046213261783123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.018176885321736336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_tensor([1077, 1077])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_tensor([1077, 1077])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1b73c885b45eb2f7ade2e5dbd8c4724473b63a82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_tensor([1077, 1077])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, epic wide shot begins, the camera slowly and majestically ascending to reveal a vast curtain of water thundering furiously over jagged, ancient rock cliffs. The sheer rock faces, slick with dark green moss and streaked with age's dampness, frame the relentless torrent. The water itself is a chaotic tapestry of explosive white foam and translucent, powerful streams plunging with immense force, crashing into a turbulent, unseen basin below. This onslaught creates a thick, roaring mist that surges and swirls, partially covering the base of the waterfall and capturing the cool, diffuse light of an overcast sky. The entire scene is rendered in a cinematic style, with desaturated colors dominated by the deep charcoal of the damp rock, the emerald of the surrounding pristine forest, and the bright, tumbling waters. A fine layer of film grain is skillfully applied to the image, adding a rich, tactile texture to the roaring mist and enhancing the raw, awe-inspiring atmosphere of untamed natural forces.", + "seed": 28823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_1077_seed_28823_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036377696925774215, + "vlm_logprob": -0.036377696925774215, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.036377696925774215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12847496569156647 + ], + "opposite_logprob_tokens": [ + -2.2534749507904053 + ], + "logprob_reward": -0.12847496569156647, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12847496569156647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24333199858665466, + 0.06732697784900665, + 0.06268680095672607, + 0.22611713409423828, + 0.056651752442121506, + 0.05612718313932419, + 0.06410287320613861, + 0.17124919593334198, + 0.03921659663319588, + 0.013189472258090973 + ], + "credit_map": [ + 0.02819465845823288, + 0.036669161170721054, + 0.035347554832696915, + 0.04354345425963402, + 0.018141619861125946, + 0.017203064635396004, + 0.049422357231378555, + 0.01481011975556612, + 0.008078290149569511, + 0.0066175456158816814, + 0.009928609244525433, + 0.01030326634645462, + 0.005930657964199781, + 0.004783510230481625, + 0.017555760219693184, + 0.004129340406507254, + 0.00895256269723177, + 0.011901699006557465, + 0.00829798448830843, + 0.00815632939338684, + 0.006694585084915161, + 0.0053611621260643005, + 0.007866790518164635, + 0.005455694627016783, + 0.024261023849248886, + 0.019481947645545006, + 0.029728572815656662, + 0.03428403660655022, + 0.0192127488553524, + 0.022559305652976036, + 0.05422987416386604, + 0.02235962636768818, + 0.007042294833809137, + 0.008021131157875061, + 0.008961466141045094, + 0.007183075416833162, + 0.005654523149132729, + 0.004178950563073158, + 0.010042098350822926, + 0.0055682179518043995, + 0.0050486354157328606, + 0.0036330725997686386, + 0.016514910385012627, + 0.003787991125136614, + 0.0037718284875154495, + 0.0023330096155405045, + 0.01837523840367794, + 0.0026624982710927725, + 0.006926867179572582, + 0.011167890392243862, + 0.00905932579189539, + 0.009112891741096973, + 0.009569345973432064, + 0.005295305047184229, + 0.007327931933104992, + 0.005643312353640795, + 0.025756806135177612, + 0.01221271138638258, + 0.038274794816970825, + 0.011862565763294697, + 0.011512975208461285, + 0.010304824449121952, + 0.04890996217727661, + 0.01241455227136612, + 0.00692810770124197, + 0.004986260086297989, + 0.010631939396262169, + 0.0041765789501369, + 0.003125646850094199, + 0.002621325431391597, + 0.0037998510524630547, + 0.0029468878637999296, + 0.00196808367036283, + 0.0015562233747914433, + 0.001765235560014844, + 0.0012877098051831126, + 0.0011128208134323359, + 0.0013591108145192266, + 0.002442635828629136, + 0.0016976530896499753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.12847496569156647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638520672917366 + ], + "opposite_logprob_tokens": [ + -3.6513853073120117 + ], + "logprob_reward": -0.02638520672917366, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638520672917366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2766086757183075, + 0.13928355276584625, + 0.0883110910654068, + 0.10052099823951721, + 0.05284387990832329, + 0.03010605089366436, + 0.0451396107673645, + 0.2010277360677719, + 0.04072333499789238, + 0.02543509751558304 + ], + "credit_map": [ + 0.031934358179569244, + 0.03574836626648903, + 0.048874907195568085, + 0.04517122730612755, + 0.03320334106683731, + 0.025058569386601448, + 0.04349145665764809, + 0.013126452453434467, + 0.013661041855812073, + 0.011776347644627094, + 0.02294093742966652, + 0.024603575468063354, + 0.008146487176418304, + 0.006216426845639944, + 0.04525727778673172, + 0.006681456696242094, + 0.013924884609878063, + 0.015934376046061516, + 0.012456952594220638, + 0.010375439189374447, + 0.009429876692593098, + 0.00828249566257, + 0.009721823036670685, + 0.008185246959328651, + 0.010768982581794262, + 0.006929901894181967, + 0.011772939935326576, + 0.018189992755651474, + 0.007348715793341398, + 0.008432416245341301, + 0.029878845438361168, + 0.007199205923825502, + 0.006412790156900883, + 0.007696065120398998, + 0.008089330047369003, + 0.006413752678781748, + 0.004817018751055002, + 0.004728798288851976, + 0.005770708434283733, + 0.008915417827665806, + 0.003054962260648608, + 0.003495260141789913, + 0.008336909115314484, + 0.0036067734472453594, + 0.0022193696349859238, + 0.0019281324930489063, + 0.005155718419700861, + 0.002308923751115799, + 0.0052942135371267796, + 0.009475314989686012, + 0.005928521044552326, + 0.005724979564547539, + 0.0057272776030004025, + 0.004564063623547554, + 0.004692872520536184, + 0.003732370911166072, + 0.030482221394777298, + 0.015372448600828648, + 0.04474824666976929, + 0.014410453848540783, + 0.01390883419662714, + 0.012104072608053684, + 0.05579051375389099, + 0.014210959896445274, + 0.005356896668672562, + 0.005625928286463022, + 0.006228611338883638, + 0.003805378684774041, + 0.007158913649618626, + 0.00398262171074748, + 0.004746277816593647, + 0.003818704281002283, + 0.0032729841768741608, + 0.004597191233187914, + 0.003018411574885249, + 0.0025290914345532656, + 0.0032663990277796984, + 0.003033056389540434, + 0.0025514434091746807, + 0.003166518872603774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02638520672917366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028254371136426926 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.0028254371136426926, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028254371136426926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18329308927059174, + 0.29379165172576904, + 0.04800412431359291, + 0.07951319217681885, + 0.0299863051623106, + 0.022669006139039993, + 0.039812762290239334, + 0.23756003379821777, + 0.037820812314748764, + 0.027549056336283684 + ], + "credit_map": [ + 0.0189285296946764, + 0.027640894055366516, + 0.023418588563799858, + 0.0343361459672451, + 0.019400035962462425, + 0.014754383824765682, + 0.033642131835222244, + 0.011172383092343807, + 0.02612997218966484, + 0.015389466658234596, + 0.0521794930100441, + 0.05708875134587288, + 0.016553044319152832, + 0.010633451864123344, + 0.10433199256658554, + 0.011485490016639233, + 0.006556352600455284, + 0.010445849038660526, + 0.006538233254104853, + 0.006311170291155577, + 0.004952071234583855, + 0.004407482221722603, + 0.004461213480681181, + 0.004331754986196756, + 0.009174841456115246, + 0.009610776789486408, + 0.011285199783742428, + 0.01257564127445221, + 0.008122327737510204, + 0.006509788334369659, + 0.016666727140545845, + 0.005567897576838732, + 0.0031187806744128466, + 0.004169792402535677, + 0.004405831452459097, + 0.004646130837500095, + 0.002743231365457177, + 0.0022906213998794556, + 0.004508622922003269, + 0.0041032941080629826, + 0.002425473416224122, + 0.0033316740300506353, + 0.0059503852389752865, + 0.002385892206802964, + 0.0016538378549739718, + 0.001082106027752161, + 0.004660099279135466, + 0.0011795389000326395, + 0.005765568930655718, + 0.010538819245994091, + 0.004640307743102312, + 0.004755584057420492, + 0.004345041234046221, + 0.0028188570868223906, + 0.003759124781936407, + 0.0031894624698907137, + 0.03669247776269913, + 0.019086336717009544, + 0.05386558547616005, + 0.01690545678138733, + 0.016185713931918144, + 0.012940574437379837, + 0.06535869836807251, + 0.01652519404888153, + 0.005953676998615265, + 0.0048002623952925205, + 0.010083138011395931, + 0.003796726232394576, + 0.0034863268956542015, + 0.002551741898059845, + 0.0038347397930920124, + 0.0033142014872282743, + 0.003311822423711419, + 0.00339525262825191, + 0.003078250912949443, + 0.002924896776676178, + 0.002714896108955145, + 0.0032185742165893316, + 0.003977391868829727, + 0.004927974194288254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0028254371136426926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40085798501968384, + 0.14720997214317322, + 0.03888074308633804, + 0.08988463133573532, + 0.05533702298998833, + 0.045664045959711075, + 0.059023596346378326, + 0.09702201932668686, + 0.0346267968416214, + 0.031493157148361206 + ], + "credit_map": [ + 0.04872797057032585, + 0.0615600049495697, + 0.048020947724580765, + 0.06003497913479805, + 0.05192588269710541, + 0.03875806927680969, + 0.06585317105054855, + 0.025976981967687607, + 0.014134163968265057, + 0.011283381842076778, + 0.034739766269922256, + 0.020375540480017662, + 0.009964033961296082, + 0.009576881304383278, + 0.040324676781892776, + 0.00681152893230319, + 0.003826283384114504, + 0.005912902764976025, + 0.003500427817925811, + 0.0066869244910776615, + 0.003987760748714209, + 0.004355692770332098, + 0.005557005759328604, + 0.005053745117038488, + 0.010053375735878944, + 0.010149256326258183, + 0.011748778633773327, + 0.015245188027620316, + 0.008141838014125824, + 0.010237548500299454, + 0.017559431493282318, + 0.006749205756932497, + 0.005710185505449772, + 0.0056004696525633335, + 0.007831203751266003, + 0.006077870726585388, + 0.0058418940752744675, + 0.005621182732284069, + 0.013149463571608067, + 0.0055047557689249516, + 0.0033128103241324425, + 0.0032636113464832306, + 0.013301163911819458, + 0.003750345902517438, + 0.003701230976730585, + 0.0025941645726561546, + 0.013289174996316433, + 0.0024515397381037474, + 0.005782519467175007, + 0.008888758718967438, + 0.01020924374461174, + 0.007991380989551544, + 0.007820559665560722, + 0.005766946356743574, + 0.006927076261490583, + 0.005637111142277718, + 0.013268525712192059, + 0.006927386857569218, + 0.02599029429256916, + 0.007852076552808285, + 0.006313068326562643, + 0.005573731381446123, + 0.024561643600463867, + 0.00653529679402709, + 0.005002094432711601, + 0.004443569108843803, + 0.008685639128088951, + 0.003194392891600728, + 0.0041467235423624516, + 0.0025646400172263384, + 0.003196870209649205, + 0.003392865415662527, + 0.004226832650601864, + 0.004186083097010851, + 0.0033809340093284845, + 0.005141947418451309, + 0.004159404896199703, + 0.0035376502200961113, + 0.0034096064046025276, + 0.0034506991505622864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0036176491994410753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585225895047188 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.020585225895047188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585225895047188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3200504779815674, + 0.24553802609443665, + 0.026603784412145615, + 0.09792790561914444, + 0.055285245180130005, + 0.04437290504574776, + 0.0499812550842762, + 0.12285061180591583, + 0.027723193168640137, + 0.00966650526970625 + ], + "credit_map": [ + 0.025639472529292107, + 0.029290754348039627, + 0.03915710374712944, + 0.07145324349403381, + 0.03703656420111656, + 0.025012914091348648, + 0.0748281329870224, + 0.017632316797971725, + 0.023521514609456062, + 0.01173432543873787, + 0.0425029993057251, + 0.04813818633556366, + 0.01331278681755066, + 0.00954436045140028, + 0.08818425983190536, + 0.008599628694355488, + 0.002439651172608137, + 0.0047602723352611065, + 0.003462031716480851, + 0.0035430919378995895, + 0.0030871815979480743, + 0.002919319784268737, + 0.003265188541263342, + 0.0031270498875528574, + 0.010235361754894257, + 0.007296286057680845, + 0.013396622613072395, + 0.015431187115609646, + 0.009714223444461823, + 0.008610074408352375, + 0.024257628247141838, + 0.00898653082549572, + 0.005387264769524336, + 0.005185267422348261, + 0.008030008524656296, + 0.0077038975432515144, + 0.005450697615742683, + 0.005097603891044855, + 0.014241623692214489, + 0.00418888870626688, + 0.004458329640328884, + 0.00496381800621748, + 0.010250035673379898, + 0.0029120107647031546, + 0.0036448915489017963, + 0.002500308910384774, + 0.01233384758234024, + 0.0033096675761044025, + 0.0068483916111290455, + 0.011423937045037746, + 0.007357224356383085, + 0.005957184359431267, + 0.004221838433295488, + 0.004181448370218277, + 0.005180140025913715, + 0.004811099264770746, + 0.016938678920269012, + 0.010248295962810516, + 0.026685774326324463, + 0.009549015201628208, + 0.01055774837732315, + 0.007327533792704344, + 0.032831110060214996, + 0.008712473325431347, + 0.0034645250998437405, + 0.0038199792616069317, + 0.004528613295406103, + 0.002846349263563752, + 0.0037656426429748535, + 0.0028498985338956118, + 0.0027434369549155235, + 0.0037047499790787697, + 0.0012710490263998508, + 0.0011973147047683597, + 0.002103277947753668, + 0.0012993128038942814, + 0.0009419020498171449, + 0.0009578330791555345, + 0.0007959430222399533, + 0.001099873916245997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020585225895047188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, epic wide shot begins, the camera slowly and majestically ascending to reveal a vast curtain of water thundering furiously over jagged, ancient rock cliffs. The sheer rock faces, slick with dark green moss and streaked with age's dampness, frame the relentless torrent. The water itself is a chaotic tapestry of explosive white foam and translucent, powerful streams plunging with immense force, crashing into a turbulent, unseen basin below. This onslaught creates a thick, roaring mist that surges and swirls, partially covering the base of the waterfall and capturing the cool, diffuse light of an overcast sky. The entire scene is rendered in a cinematic style, with desaturated colors dominated by the deep charcoal of the damp rock, the emerald of the surrounding pristine forest, and the bright, tumbling waters. A fine layer of film grain is skillfully applied to the image, adding a rich, tactile texture to the roaring mist and enhancing the raw, awe-inspiring atmosphere of untamed natural forces.", + "seed": 28824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_1077_seed_28824_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010701001714915037, + "vlm_logprob": -0.010701001714915037, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.010701001714915037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02922585979104042 + ], + "opposite_logprob_tokens": [ + -3.9042258262634277 + ], + "logprob_reward": -0.02922585979104042, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02922585979104042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.135394886136055, + 0.07859068363904953, + 0.06831882148981094, + 0.04199989140033722, + 0.10576193779706955, + 0.05033281818032265, + 0.10445161908864975, + 0.016232965514063835, + 0.04376837983727455, + 0.35514798760414124 + ], + "credit_map": [ + 0.013820289634168148, + 0.022505326196551323, + 0.02050693705677986, + 0.015348787419497967, + 0.015317391604185104, + 0.012146150693297386, + 0.016870055347681046, + 0.018879974260926247, + 0.0063762906938791275, + 0.015531095676124096, + 0.00854793842881918, + 0.013827778398990631, + 0.011776578612625599, + 0.007337395567446947, + 0.006738858763128519, + 0.008454750292003155, + 0.007117080502212048, + 0.010541418567299843, + 0.007069343235343695, + 0.00944574922323227, + 0.00956111028790474, + 0.009201745502650738, + 0.00968430656939745, + 0.005698076449334621, + 0.003531844587996602, + 0.007921602576971054, + 0.005937687586992979, + 0.006359946448355913, + 0.005067519377917051, + 0.002929558977484703, + 0.00721399113535881, + 0.0030377404764294624, + 0.008478706702589989, + 0.013253580778837204, + 0.016189182177186012, + 0.016974691301584244, + 0.011547453701496124, + 0.008251328021287918, + 0.02123592235147953, + 0.009831083938479424, + 0.006925414316356182, + 0.008465306833386421, + 0.005087722558528185, + 0.008455074392259121, + 0.004702094942331314, + 0.005059370305389166, + 0.006837277673184872, + 0.004800564609467983, + 0.007353895343840122, + 0.03079071454703808, + 0.006889133248478174, + 0.03425385430455208, + 0.0064441608265042305, + 0.00544242188334465, + 0.0074255624786019325, + 0.0058518885634839535, + 0.001700650085695088, + 0.002020105719566345, + 0.0033896618988364935, + 0.0020042951218783855, + 0.0018218194600194693, + 0.0015276264166459441, + 0.002351433504372835, + 0.0014173752861097455, + 0.00333936489187181, + 0.010651253163814545, + 0.004158808384090662, + 0.011638284660875797, + 0.0033316973131150007, + 0.0025733080692589283, + 0.005595965776592493, + 0.0024797003716230392, + 0.006794852204620838, + 0.14581336081027985, + 0.009570091031491756, + 0.154375821352005, + 0.01449630782008171, + 0.004002754110842943, + 0.012916089966893196, + 0.007178757805377245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.02922585979104042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00465749204158783 + ], + "opposite_logprob_tokens": [ + -5.37965726852417 + ], + "logprob_reward": -0.00465749204158783, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00465749204158783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055839281529188156, + 0.023052111268043518, + 0.04264497756958008, + 0.0198465995490551, + 0.01923227123916149, + 0.01740438863635063, + 0.024559753015637398, + 0.016068758442997932, + 0.10745116323232651, + 0.6739007830619812 + ], + "credit_map": [ + 0.0041114334017038345, + 0.008610600605607033, + 0.005198969040066004, + 0.005137058440595865, + 0.013347547501325607, + 0.004261639900505543, + 0.010553077794611454, + 0.004618946462869644, + 0.0020712141413241625, + 0.003819771111011505, + 0.002459472045302391, + 0.003105123993009329, + 0.0043229456059634686, + 0.0023707221262156963, + 0.003011972876265645, + 0.0018908872734755278, + 0.0017079542158171535, + 0.013359076343476772, + 0.0023524616844952106, + 0.015635017305612564, + 0.003045459045097232, + 0.0016497422475367785, + 0.002910194918513298, + 0.0019850668031722307, + 0.001552106929011643, + 0.004353745374828577, + 0.002258549677208066, + 0.0032948011066764593, + 0.0025160200893878937, + 0.001294598332606256, + 0.003028232604265213, + 0.0015485434560105205, + 0.0015260172076523304, + 0.002996914554387331, + 0.0025268211029469967, + 0.0031234777998179197, + 0.0024612832348793745, + 0.0014542242279276252, + 0.0033694645389914513, + 0.0017740665934979916, + 0.002087828703224659, + 0.0025416517164558172, + 0.002345083514228463, + 0.002484450116753578, + 0.0018549554515630007, + 0.0016862826887518167, + 0.0026974258944392204, + 0.0017067079897969961, + 0.002283193403854966, + 0.0037940521724522114, + 0.0038657584227621555, + 0.004100334830582142, + 0.0026501112151890993, + 0.0024501443840563297, + 0.003069226862862706, + 0.00234692869707942, + 0.0015946835046634078, + 0.002026960486546159, + 0.0038409738335758448, + 0.002621247200295329, + 0.0016255867667496204, + 0.0012509272200986743, + 0.0017770177219063044, + 0.0013313591480255127, + 0.003746342845261097, + 0.03951212391257286, + 0.003638742258772254, + 0.04127536714076996, + 0.00578899635002017, + 0.002710554748773575, + 0.007773774676024914, + 0.0030052550137043, + 0.011291341856122017, + 0.28015339374542236, + 0.01772531494498253, + 0.2972657084465027, + 0.02677847631275654, + 0.005247877910733223, + 0.022753693163394928, + 0.012684891931712627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.00465749204158783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07183302193880081, + 0.08152244985103607, + 0.09132594615221024, + 0.023789040744304657, + 0.02856292389333248, + 0.03680895268917084, + 0.09085366874933243, + 0.04726541042327881, + 0.2534804046154022, + 0.2745581865310669 + ], + "credit_map": [ + 0.006105158943682909, + 0.009093132801353931, + 0.010286235250532627, + 0.008602337911725044, + 0.010991138406097889, + 0.006616304628551006, + 0.00959018338471651, + 0.010548536665737629, + 0.007198280654847622, + 0.014512156136333942, + 0.0077134789898991585, + 0.01406134944409132, + 0.012944403104484081, + 0.0073863789439201355, + 0.011466104537248611, + 0.006240298971533775, + 0.004606090486049652, + 0.03046220727264881, + 0.004716136492788792, + 0.026064598932862282, + 0.006903748959302902, + 0.007985126227140427, + 0.005287330597639084, + 0.005300708580762148, + 0.0020372774451971054, + 0.004248116631060839, + 0.0029172145295888186, + 0.0034293923527002335, + 0.003367988858371973, + 0.0024505152832716703, + 0.0030721912626177073, + 0.002266347873955965, + 0.002821295754984021, + 0.0035112923942506313, + 0.0031842871103435755, + 0.003639108967036009, + 0.003736636834219098, + 0.00541380699723959, + 0.0025045291986316442, + 0.0037519687321037054, + 0.00605332525447011, + 0.006155849900096655, + 0.003981900401413441, + 0.006213825661689043, + 0.002786473371088505, + 0.00396509375423193, + 0.003948405385017395, + 0.00370407965965569, + 0.005084260366857052, + 0.029858076944947243, + 0.00406683562323451, + 0.03129873424768448, + 0.0059622288681566715, + 0.004186453763395548, + 0.006368902511894703, + 0.004028188064694405, + 0.0040261694230139256, + 0.009492934681475163, + 0.0065633272752165794, + 0.008169417269527912, + 0.00533889839425683, + 0.004892819095402956, + 0.004657672252506018, + 0.004124173428863287, + 0.004792281426489353, + 0.10264793783426285, + 0.007181023713201284, + 0.10697722434997559, + 0.010215850546956062, + 0.00501348776742816, + 0.010992658324539661, + 0.005659950897097588, + 0.008889404125511646, + 0.10561046749353409, + 0.009107021614909172, + 0.11520963907241821, + 0.010297750122845173, + 0.006627656519412994, + 0.01106037013232708, + 0.007755895610898733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627884857356548 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627884857356548, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627884857356548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1225813552737236, + 0.02892528846859932, + 0.10568267852067947, + 0.04354076832532883, + 0.04448266327381134, + 0.033572614192962646, + 0.045730866491794586, + 0.03448427841067314, + 0.2534392178058624, + 0.28756025433540344 + ], + "credit_map": [ + 0.011686548590660095, + 0.02183680422604084, + 0.009867061860859394, + 0.009057927876710892, + 0.02560671605169773, + 0.009964721277356148, + 0.024354970082640648, + 0.010206588543951511, + 0.003452758537605405, + 0.003926778677850962, + 0.0033862385898828506, + 0.0039686900563538074, + 0.005138779524713755, + 0.0023819012567400932, + 0.00357506750151515, + 0.003095074789598584, + 0.004034021403640509, + 0.03727196156978607, + 0.004173473920673132, + 0.039769504219293594, + 0.006581794936209917, + 0.0034149184357374907, + 0.006166232284158468, + 0.004270766396075487, + 0.0024122209288179874, + 0.007451984565705061, + 0.003891471540555358, + 0.00846259854733944, + 0.0043280040845274925, + 0.0028684886638075113, + 0.010857143439352512, + 0.003268852597102523, + 0.00362684135325253, + 0.007468069437891245, + 0.004450202453881502, + 0.007113017141819, + 0.005930233746767044, + 0.003449486568570137, + 0.008562534116208553, + 0.0038822751957923174, + 0.002382232341915369, + 0.006862203124910593, + 0.003663593903183937, + 0.007906950078904629, + 0.0030768990982323885, + 0.0020465459674596786, + 0.004365870263427496, + 0.0032683159224689007, + 0.004275723826140165, + 0.010781086049973965, + 0.0043252985924482346, + 0.012303515337407589, + 0.004251029808074236, + 0.001892224419862032, + 0.004999145865440369, + 0.0029028344433754683, + 0.0028557979967445135, + 0.00934992078691721, + 0.0034421109594404697, + 0.008398457430303097, + 0.0036521952133625746, + 0.0018512688111513853, + 0.0025976456236094236, + 0.002336878329515457, + 0.006143386475741863, + 0.1006307303905487, + 0.0084381690248847, + 0.10534893721342087, + 0.010033870115876198, + 0.004418304190039635, + 0.012385704554617405, + 0.006040120031684637, + 0.005079415626823902, + 0.11881940811872482, + 0.007806395646184683, + 0.12560433149337769, + 0.01201049704104662, + 0.002923107473179698, + 0.009611889719963074, + 0.005705196410417557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.007627884857356548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252964794635773 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252964794635773, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252964794635773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09143365919589996, + 0.20488961040973663, + 0.037810441106557846, + 0.016848834231495857, + 0.03766283765435219, + 0.025353800505399704, + 0.05190228670835495, + 0.039109811186790466, + 0.05825776606798172, + 0.4367309510707855 + ], + "credit_map": [ + 0.011180967092514038, + 0.012063603848218918, + 0.008017751388251781, + 0.010203579440712929, + 0.01831716299057007, + 0.00718471547588706, + 0.018666263669729233, + 0.005799619480967522, + 0.01242329366505146, + 0.05771953612565994, + 0.012690965086221695, + 0.05026250332593918, + 0.021295534446835518, + 0.012264806777238846, + 0.021534545347094536, + 0.016698431223630905, + 0.0028023882769048214, + 0.0075873867608606815, + 0.003096607280895114, + 0.007633598987013102, + 0.004362933337688446, + 0.005776943638920784, + 0.003643933916464448, + 0.0029066484421491623, + 0.0011346355313435197, + 0.003342561423778534, + 0.0015091169625520706, + 0.0034131892025470734, + 0.002243414055556059, + 0.0010267384350299835, + 0.0029555729124695063, + 0.0012236052425578237, + 0.003451487747952342, + 0.005046939942985773, + 0.004410381894558668, + 0.005555546842515469, + 0.004535956773906946, + 0.004284324124455452, + 0.006538572255522013, + 0.0038396290037781, + 0.003273799316957593, + 0.004559936933219433, + 0.0024981796741485596, + 0.004213843494653702, + 0.00259615620598197, + 0.002625570632517338, + 0.003083257470279932, + 0.0025030572433024645, + 0.004115133546292782, + 0.010988978669047356, + 0.0054212030954658985, + 0.009695486165583134, + 0.004274980165064335, + 0.005734215024858713, + 0.007520252373069525, + 0.004152038134634495, + 0.003717679064720869, + 0.007048396859318018, + 0.007467716000974178, + 0.0066401101648807526, + 0.004268363583832979, + 0.0026924442499876022, + 0.004353404976427555, + 0.0029216997791081667, + 0.003795670811086893, + 0.016617590561509132, + 0.00471918098628521, + 0.017466790974140167, + 0.004132397472858429, + 0.004193012602627277, + 0.004295067861676216, + 0.0030380538664758205, + 0.006385911721736193, + 0.1806531399488449, + 0.01271835807710886, + 0.19098259508609772, + 0.01832306943833828, + 0.004456543363630772, + 0.014581141993403435, + 0.008630208671092987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.005252964794635773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, epic wide shot begins, the camera slowly and majestically ascending to reveal a vast curtain of water thundering furiously over jagged, ancient rock cliffs. The sheer rock faces, slick with dark green moss and streaked with age's dampness, frame the relentless torrent. The water itself is a chaotic tapestry of explosive white foam and translucent, powerful streams plunging with immense force, crashing into a turbulent, unseen basin below. This onslaught creates a thick, roaring mist that surges and swirls, partially covering the base of the waterfall and capturing the cool, diffuse light of an overcast sky. The entire scene is rendered in a cinematic style, with desaturated colors dominated by the deep charcoal of the damp rock, the emerald of the surrounding pristine forest, and the bright, tumbling waters. A fine layer of film grain is skillfully applied to the image, adding a rich, tactile texture to the roaring mist and enhancing the raw, awe-inspiring atmosphere of untamed natural forces.", + "seed": 28825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_1077_seed_28825_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01952976672910154, + "vlm_logprob": -0.01952976672910154, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01952976672910154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05049528554081917 + ], + "opposite_logprob_tokens": [ + -3.3004953861236572 + ], + "logprob_reward": -0.05049528554081917, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05049528554081917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06444451957941055, + 0.02801213599741459, + 0.013868910260498524, + 0.018450262024998665, + 0.012486009858548641, + 0.01588142104446888, + 0.01877092756330967, + 0.799929678440094, + 0.014495362527668476, + 0.013660892844200134 + ], + "credit_map": [ + 0.0072260042652487755, + 0.011363143101334572, + 0.008735914714634418, + 0.011596554890275002, + 0.005520465783774853, + 0.005046234466135502, + 0.006735128816217184, + 0.008221074938774109, + 0.00422222726047039, + 0.005510265007615089, + 0.0034046873915940523, + 0.0039764633402228355, + 0.002484790747985244, + 0.002330170478671789, + 0.002638507168740034, + 0.0034450239036232233, + 0.001255309209227562, + 0.0019484458025544882, + 0.0017632829258218408, + 0.0015541521133854985, + 0.0015195540618151426, + 0.001547527965158224, + 0.0013472354039549828, + 0.002933403244242072, + 0.0017889162991195917, + 0.003239872632548213, + 0.0034083155915141106, + 0.002499152207747102, + 0.0010855073342099786, + 0.0012312616454437375, + 0.0030322205275297165, + 0.0021650155540555716, + 0.0017990842461585999, + 0.0023573972284793854, + 0.0015884126769378781, + 0.0016438581515103579, + 0.001120653934776783, + 0.0009566254448145628, + 0.0011178983841091394, + 0.0019020793261006474, + 0.0027186633087694645, + 0.003317417809739709, + 0.002315904712304473, + 0.0023732720874249935, + 0.0009279248188249767, + 0.0013531052973121405, + 0.0012250177096575499, + 0.001650114543735981, + 0.0015238805208355188, + 0.0022234530188143253, + 0.0043507469817996025, + 0.0028662688564509153, + 0.002704975660890341, + 0.001496650162152946, + 0.002128993161022663, + 0.0014759596670046449, + 0.07030463218688965, + 0.029529765248298645, + 0.2412061095237732, + 0.065823033452034, + 0.03525293990969658, + 0.031067077070474625, + 0.27586546540260315, + 0.050880588591098785, + 0.0022797698620706797, + 0.0029067888390272856, + 0.002146507380530238, + 0.0019396557472646236, + 0.0009483056492172182, + 0.00140916567761451, + 0.0010318085551261902, + 0.0018333601765334606, + 0.002281045075505972, + 0.002401358913630247, + 0.0026375246234238148, + 0.0017868103459477425, + 0.0011108431499451399, + 0.0010941235814243555, + 0.0009249605936929584, + 0.0014242277247831225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.05049528554081917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370346546173096 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370346546173096, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370346546173096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026004303246736526, + 0.012647008523344994, + 0.0110520264133811, + 0.009997652843594551, + 0.008443607948720455, + 0.00990477204322815, + 0.010736845433712006, + 0.8876071572303772, + 0.01546595897525549, + 0.008140657097101212 + ], + "credit_map": [ + 0.0035897823981940746, + 0.0038778900634497404, + 0.0037682978436350822, + 0.0034272039774805307, + 0.0030821426771581173, + 0.0019634810741990805, + 0.0027028124313801527, + 0.0035926930140703917, + 0.001639559748582542, + 0.002656088210642338, + 0.0014104729052633047, + 0.0018923843745142221, + 0.001239831792190671, + 0.0010249497136101127, + 0.001244388404302299, + 0.001539334305562079, + 0.0012783777201548219, + 0.0027472416404634714, + 0.0015353638445958495, + 0.0013169938465580344, + 0.0007755974074825644, + 0.0007747505442239344, + 0.0010810804087668657, + 0.0015426207100972533, + 0.0008985029999166727, + 0.0015794928185641766, + 0.002030997071415186, + 0.0012751654721796513, + 0.0006977487937547266, + 0.0006400272250175476, + 0.0018472285009920597, + 0.0010284894378855824, + 0.0008689634269103408, + 0.0016722689615562558, + 0.0009575991425663233, + 0.0012960637686774135, + 0.0006720277015119791, + 0.000790154212154448, + 0.0009093183325603604, + 0.0012772131012752652, + 0.0016550839645788074, + 0.0011457165237516165, + 0.0011946165468543768, + 0.0013895239681005478, + 0.001352539169602096, + 0.0008770555723458529, + 0.0013232644414529204, + 0.0009669715072959661, + 0.0009200694039463997, + 0.0010543301468715072, + 0.001680443063378334, + 0.0014023617841303349, + 0.0018355570500716567, + 0.001080355723388493, + 0.0013257553800940514, + 0.0014379731146618724, + 0.07802373170852661, + 0.0327179953455925, + 0.267559289932251, + 0.07332629710435867, + 0.03896558657288551, + 0.034430354833602905, + 0.30637645721435547, + 0.056207455694675446, + 0.0018081567250192165, + 0.0013856309233233333, + 0.004437175113707781, + 0.002242456190288067, + 0.0010625645518302917, + 0.0013705146266147494, + 0.001822222606278956, + 0.0013372383546084166, + 0.0010905207600444555, + 0.0013863870408385992, + 0.0011272861156612635, + 0.0010307710617780685, + 0.000954927469138056, + 0.0008605789043940604, + 0.0007298369891941547, + 0.0009603481739759445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.026370346546173096 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645507782697678 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645507782697678, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645507782697678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12224384397268295, + 0.0622117817401886, + 0.046198248863220215, + 0.06850486248731613, + 0.036075443029403687, + 0.06472183763980865, + 0.0408439040184021, + 0.5153592228889465, + 0.02034895308315754, + 0.02349187061190605 + ], + "credit_map": [ + 0.015229873359203339, + 0.020964596420526505, + 0.016154754906892776, + 0.020259395241737366, + 0.010951383039355278, + 0.0127538051456213, + 0.011985366232693195, + 0.013944664038717747, + 0.006531378719955683, + 0.013488245196640491, + 0.007535511162132025, + 0.009066500701010227, + 0.005708078853785992, + 0.007304713129997253, + 0.006020368076860905, + 0.006556989625096321, + 0.006368427071720362, + 0.008417338132858276, + 0.0054703205823898315, + 0.005668839439749718, + 0.00428105378523469, + 0.004223281517624855, + 0.004241925198584795, + 0.007527064997702837, + 0.007504752837121487, + 0.011402410455048084, + 0.014544540084898472, + 0.005392909049987793, + 0.004772982560098171, + 0.0046281092800199986, + 0.012459103018045425, + 0.0078000533394515514, + 0.004177940543740988, + 0.007594936527311802, + 0.0039035496301949024, + 0.004880087450146675, + 0.0037609711289405823, + 0.0035144835710525513, + 0.003578625852242112, + 0.0046648504212498665, + 0.010192401707172394, + 0.011398056522011757, + 0.00801725685596466, + 0.008968759328126907, + 0.006303503178060055, + 0.00656428188085556, + 0.006776314694434404, + 0.006501261610537767, + 0.004990645218640566, + 0.0037383027374744415, + 0.011996643617749214, + 0.007150489836931229, + 0.0025345843750983477, + 0.003300276817753911, + 0.0031645046547055244, + 0.003968456294387579, + 0.04476933181285858, + 0.019181007519364357, + 0.15657958388328552, + 0.04233451187610626, + 0.0226709246635437, + 0.020272336900234222, + 0.17690838873386383, + 0.0326431468129158, + 0.002863279776647687, + 0.004155072383582592, + 0.0031315642409026623, + 0.0020315570291131735, + 0.0018916784320026636, + 0.0020561697892844677, + 0.001467728870920837, + 0.0027519031427800655, + 0.0027361498214304447, + 0.00344741134904325, + 0.003991995006799698, + 0.0032952704932540655, + 0.002391453832387924, + 0.002279229462146759, + 0.002305207774043083, + 0.0030451519414782524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.004645507782697678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019406310748308897 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019406310748308897, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019406310748308897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11525582522153854, + 0.054869651794433594, + 0.05074445530772209, + 0.05591197311878204, + 0.10155486315488815, + 0.07930117845535278, + 0.07163837552070618, + 0.4225945472717285, + 0.03222141042351723, + 0.01590769737958908 + ], + "credit_map": [ + 0.013954225927591324, + 0.017168885096907616, + 0.022571062669157982, + 0.013893742114305496, + 0.010867555625736713, + 0.009242466650903225, + 0.017127977684140205, + 0.010429898276925087, + 0.004921713378280401, + 0.007575977127999067, + 0.006857900880277157, + 0.008082319051027298, + 0.00553244398906827, + 0.006359049119055271, + 0.00701593654230237, + 0.00852430984377861, + 0.007267146836966276, + 0.008829877711832523, + 0.008523697964847088, + 0.007353015709668398, + 0.0033516008406877518, + 0.004320119507610798, + 0.006160201039165258, + 0.004938795696943998, + 0.008085566572844982, + 0.010095570236444473, + 0.006567088421434164, + 0.008686297573149204, + 0.004896456375718117, + 0.005470734555274248, + 0.005525479093194008, + 0.006584781222045422, + 0.018021423369646072, + 0.009182439185678959, + 0.019335942342877388, + 0.013137987814843655, + 0.005457382649183273, + 0.005602729972451925, + 0.007425243500620127, + 0.023391718044877052, + 0.00974371936172247, + 0.006654647178947926, + 0.011818400584161282, + 0.011903856880962849, + 0.006867577787488699, + 0.008772963657975197, + 0.010068447329103947, + 0.013471565209329128, + 0.006127197295427322, + 0.005942706950008869, + 0.017608851194381714, + 0.010724780149757862, + 0.0071402317844331264, + 0.006510953418910503, + 0.006897363346070051, + 0.010686291381716728, + 0.03649744763970375, + 0.021948369219899178, + 0.1143251582980156, + 0.036655910313129425, + 0.0226128026843071, + 0.021672548726201057, + 0.13812962174415588, + 0.030752696096897125, + 0.0053412276320159435, + 0.0037777815014123917, + 0.006788699887692928, + 0.004063057247549295, + 0.0025821903254836798, + 0.0029508285224437714, + 0.0034692897461354733, + 0.0032483353279531, + 0.002395134186372161, + 0.0022963332012295723, + 0.0025120105128735304, + 0.002549736760556698, + 0.0012020576978102326, + 0.0014864159747958183, + 0.0013962468365207314, + 0.0020697612781077623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0019406310748308897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15983614325523376, + 0.09306173771619797, + 0.04831283167004585, + 0.06101807579398155, + 0.055867403745651245, + 0.047188904136419296, + 0.028007229790091515, + 0.46859830617904663, + 0.01903572306036949, + 0.019073674455285072 + ], + "credit_map": [ + 0.015485802665352821, + 0.018043071031570435, + 0.028186867013573647, + 0.029857676476240158, + 0.011143390089273453, + 0.012509000487625599, + 0.022445132955908775, + 0.022165188565850258, + 0.012025784701108932, + 0.013151378370821476, + 0.011304412968456745, + 0.01464642584323883, + 0.01100247073918581, + 0.008668225258588791, + 0.011735755018889904, + 0.010527273640036583, + 0.005511756520718336, + 0.008454401977360249, + 0.008540412411093712, + 0.00489024305716157, + 0.004256513901054859, + 0.00430480670183897, + 0.004335265606641769, + 0.00801942404359579, + 0.006470969412475824, + 0.007647351827472448, + 0.014780514873564243, + 0.004652173258364201, + 0.004301033914089203, + 0.0032613594084978104, + 0.01300218515098095, + 0.006902483291924, + 0.009785329923033714, + 0.0037888719234615564, + 0.010045668110251427, + 0.007918774150311947, + 0.0027173866983503103, + 0.0037332067731767893, + 0.004438739735633135, + 0.013439422473311424, + 0.005961681250482798, + 0.004360937979072332, + 0.007304653525352478, + 0.008640366606414318, + 0.004514290019869804, + 0.004344767425209284, + 0.005488963332027197, + 0.0065732356160879135, + 0.0021605303045362234, + 0.0035486300475895405, + 0.004632160067558289, + 0.0027612755075097084, + 0.004525061696767807, + 0.003309206571429968, + 0.004047422669827938, + 0.0030229417607188225, + 0.04178142547607422, + 0.018659302964806557, + 0.13385452330112457, + 0.039276931434869766, + 0.021184712648391724, + 0.01898331195116043, + 0.16416682302951813, + 0.030691232532262802, + 0.003532032947987318, + 0.0031797231640666723, + 0.0024834563955664635, + 0.002450301544740796, + 0.0016259262338280678, + 0.0018197380704805255, + 0.0014786312822252512, + 0.0024659116752445698, + 0.0023548556491732597, + 0.0020821697544306517, + 0.004861945286393166, + 0.002429628511890769, + 0.0013507807161659002, + 0.002026671776548028, + 0.0015586953377351165, + 0.0024089268408715725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.014197062700986862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, epic wide shot begins, the camera slowly and majestically ascending to reveal a vast curtain of water thundering furiously over jagged, ancient rock cliffs. The sheer rock faces, slick with dark green moss and streaked with age's dampness, frame the relentless torrent. The water itself is a chaotic tapestry of explosive white foam and translucent, powerful streams plunging with immense force, crashing into a turbulent, unseen basin below. This onslaught creates a thick, roaring mist that surges and swirls, partially covering the base of the waterfall and capturing the cool, diffuse light of an overcast sky. The entire scene is rendered in a cinematic style, with desaturated colors dominated by the deep charcoal of the damp rock, the emerald of the surrounding pristine forest, and the bright, tumbling waters. A fine layer of film grain is skillfully applied to the image, adding a rich, tactile texture to the roaring mist and enhancing the raw, awe-inspiring atmosphere of untamed natural forces.", + "seed": 28826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_5/index_1077_seed_28826_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021276503521949052, + "vlm_logprob": -0.021276503521949052, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.021276503521949052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04473355412483215 + ], + "opposite_logprob_tokens": [ + -3.4197335243225098 + ], + "logprob_reward": -0.04473355412483215, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04473355412483215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10436226427555084, + 0.10988056659698486, + 0.132880300283432, + 0.04305077716708183, + 0.260050505399704, + 0.09819333255290985, + 0.08427618443965912, + 0.07678698748350143, + 0.03494767099618912, + 0.055571429431438446 + ], + "credit_map": [ + 0.01460131723433733, + 0.0176816675812006, + 0.020045962184667587, + 0.011537326499819756, + 0.009416772052645683, + 0.008520043455064297, + 0.013523728586733341, + 0.009035450406372547, + 0.014816170558333397, + 0.01993621699512005, + 0.02158375270664692, + 0.01444596890360117, + 0.010051813907921314, + 0.007357134483754635, + 0.013202025555074215, + 0.008487479761242867, + 0.012359184212982655, + 0.008521175011992455, + 0.04892077296972275, + 0.041208382695913315, + 0.004286948591470718, + 0.005263073369860649, + 0.00716147106140852, + 0.005159297026693821, + 0.005765837151557207, + 0.0075846100226044655, + 0.005489978939294815, + 0.004627897869795561, + 0.004136596340686083, + 0.0044226329773664474, + 0.005449940450489521, + 0.005573282018303871, + 0.03582776337862015, + 0.033065833151340485, + 0.08289393782615662, + 0.05322917550802231, + 0.013043363578617573, + 0.012074755504727364, + 0.02076585777103901, + 0.009149824269115925, + 0.012355822138488293, + 0.0190863199532032, + 0.012267466634511948, + 0.01558617502450943, + 0.011188318021595478, + 0.008979222737252712, + 0.01094901654869318, + 0.007780980318784714, + 0.010982437059283257, + 0.012740775011479855, + 0.012202661484479904, + 0.009890573099255562, + 0.011848094873130322, + 0.009874963201582432, + 0.009133516810834408, + 0.007603159174323082, + 0.00826420821249485, + 0.019232027232646942, + 0.010052023455500603, + 0.011242854408919811, + 0.0070446631871163845, + 0.005753191187977791, + 0.00986875593662262, + 0.0053292615339159966, + 0.003268167842179537, + 0.005186277907341719, + 0.005283264443278313, + 0.004759178962558508, + 0.003034125082194805, + 0.005336269736289978, + 0.004357175901532173, + 0.003723211120814085, + 0.003662349423393607, + 0.0038223182782530785, + 0.014669647440314293, + 0.018524225801229477, + 0.0023205564357340336, + 0.0031519047915935516, + 0.006779614370316267, + 0.0026408147532492876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.04473355412483215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331276424229145 + ], + "opposite_logprob_tokens": [ + -3.77331280708313 + ], + "logprob_reward": -0.02331276424229145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331276424229145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1277129352092743, + 0.08642377704381943, + 0.12337497621774673, + 0.062124043703079224, + 0.1480892300605774, + 0.18055011332035065, + 0.11779201775789261, + 0.08079905807971954, + 0.0387938991189003, + 0.03433996066451073 + ], + "credit_map": [ + 0.02101808600127697, + 0.022339247167110443, + 0.023032385855913162, + 0.012727834284305573, + 0.012823879718780518, + 0.010835937224328518, + 0.013255481608211994, + 0.011680081486701965, + 0.011763392947614193, + 0.017543258145451546, + 0.018476903438568115, + 0.01161924283951521, + 0.00852182786911726, + 0.00488906167447567, + 0.008104445412755013, + 0.005505648907274008, + 0.01166172232478857, + 0.011576258577406406, + 0.041279733180999756, + 0.035750340670347214, + 0.0048224409110844135, + 0.005402482580393553, + 0.0075341020710766315, + 0.005347884725779295, + 0.009736503474414349, + 0.012939824722707272, + 0.010657050646841526, + 0.008704724721610546, + 0.005175327882170677, + 0.004446833394467831, + 0.006527917459607124, + 0.003935860004276037, + 0.021255973726511, + 0.022421255707740784, + 0.038146939128637314, + 0.030054215341806412, + 0.010606152005493641, + 0.007489179261028767, + 0.011159050278365612, + 0.006956466939300299, + 0.022433148697018623, + 0.035390254110097885, + 0.022026732563972473, + 0.026363549754023552, + 0.021074345335364342, + 0.018775546923279762, + 0.01998908631503582, + 0.014497442170977592, + 0.016747966408729553, + 0.020327448844909668, + 0.018057391047477722, + 0.014537432231009007, + 0.014997717924416065, + 0.011319983750581741, + 0.012084795162081718, + 0.009719286113977432, + 0.009202543646097183, + 0.014595960266888142, + 0.014280740171670914, + 0.010454815812408924, + 0.00852615013718605, + 0.0085851876065135, + 0.008263325318694115, + 0.006890332791954279, + 0.0031426073983311653, + 0.005478651728481054, + 0.006587835028767586, + 0.007610857021063566, + 0.003275951137766242, + 0.005179135594516993, + 0.00392190832644701, + 0.003596954047679901, + 0.0044408394023776054, + 0.007460135035216808, + 0.006163313984870911, + 0.0038665966130793095, + 0.0029739737510681152, + 0.0030874169897288084, + 0.003710217773914337, + 0.002637465251609683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.02331276424229145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195966258645058 + ], + "opposite_logprob_tokens": [ + -4.018196105957031 + ], + "logprob_reward": -0.018195966258645058, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195966258645058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11193706095218658, + 0.12911084294319153, + 0.1491321325302124, + 0.023479269817471504, + 0.10885567218065262, + 0.06146626174449921, + 0.07085882127285004, + 0.15343064069747925, + 0.05621655285358429, + 0.1355127990245819 + ], + "credit_map": [ + 0.01245700754225254, + 0.01471735630184412, + 0.022404715418815613, + 0.017375975847244263, + 0.01171836256980896, + 0.011206451803445816, + 0.012654069811105728, + 0.009403111413121223, + 0.014820742420852184, + 0.014816815964877605, + 0.03568112850189209, + 0.02557680569589138, + 0.008927149698138237, + 0.00825924426317215, + 0.011095326393842697, + 0.009933620691299438, + 0.013171003200113773, + 0.015338069759309292, + 0.046336956322193146, + 0.04627595096826553, + 0.00648622028529644, + 0.006219509057700634, + 0.009338296949863434, + 0.005966102704405785, + 0.003422603476792574, + 0.00353439268656075, + 0.0049321758560836315, + 0.0034761091228574514, + 0.002302675973623991, + 0.0019111509900540113, + 0.0021121723111718893, + 0.0017879860242828727, + 0.013681585900485516, + 0.01395281869918108, + 0.03608701750636101, + 0.024494526907801628, + 0.005243451800197363, + 0.0043580904603004456, + 0.008019070141017437, + 0.003019095864146948, + 0.0075654746033251286, + 0.012241300195455551, + 0.008087147027254105, + 0.008128814399242401, + 0.006582311820238829, + 0.006547201424837112, + 0.006832504644989967, + 0.0054814997129142284, + 0.009319799952208996, + 0.01633628085255623, + 0.01235019788146019, + 0.008762400597333908, + 0.005130897741764784, + 0.0055573792196810246, + 0.0077411700040102005, + 0.005660686641931534, + 0.015736015513539314, + 0.035652391612529755, + 0.03348119184374809, + 0.0223789531737566, + 0.01151847094297409, + 0.010906835086643696, + 0.013995030894875526, + 0.009761742316186428, + 0.007692054379731417, + 0.00888020545244217, + 0.009025178849697113, + 0.010035979561507702, + 0.004190582782030106, + 0.005864656995981932, + 0.005856327246874571, + 0.004671563860028982, + 0.010119580663740635, + 0.010006660595536232, + 0.031612105667591095, + 0.04443911090493202, + 0.00553450733423233, + 0.009283808059990406, + 0.0179282259196043, + 0.00658878730610013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.018195966258645058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14577360451221466, + 0.09492025524377823, + 0.2859606444835663, + 0.06800246983766556, + 0.143995001912117, + 0.054954901337623596, + 0.0460595041513443, + 0.06769037246704102, + 0.04401117190718651, + 0.04863206669688225 + ], + "credit_map": [ + 0.022754743695259094, + 0.018151307478547096, + 0.04026905447244644, + 0.021528437733650208, + 0.010036764666438103, + 0.008505775593221188, + 0.016069995239377022, + 0.008457518182694912, + 0.01045981701463461, + 0.015081437304615974, + 0.02459230087697506, + 0.015746833756566048, + 0.008891982957720757, + 0.005784180946648121, + 0.008256795816123486, + 0.006106906570494175, + 0.025367358699440956, + 0.012593690305948257, + 0.11171332746744156, + 0.09752576053142548, + 0.008857110515236855, + 0.00867998506873846, + 0.014348812401294708, + 0.00687460508197546, + 0.008255616761744022, + 0.007835878059267998, + 0.016778111457824707, + 0.010558406822383404, + 0.007493853569030762, + 0.005391107406467199, + 0.006764869671314955, + 0.004924621898680925, + 0.02415398508310318, + 0.019518351182341576, + 0.03735225275158882, + 0.027574747800827026, + 0.008016441017389297, + 0.007372936699539423, + 0.011242980137467384, + 0.00876329094171524, + 0.007399475667625666, + 0.0071251061744987965, + 0.011068898253142834, + 0.006904704496264458, + 0.00602995278313756, + 0.0053981333039700985, + 0.00527389720082283, + 0.005754727404564619, + 0.005214371718466282, + 0.010352677665650845, + 0.006990506313741207, + 0.004577931482344866, + 0.003938465379178524, + 0.005379373673349619, + 0.005299823824316263, + 0.004306351765990257, + 0.007501149084419012, + 0.015667401254177094, + 0.013087503612041473, + 0.008480872958898544, + 0.005622992757707834, + 0.006035860627889633, + 0.006202692165970802, + 0.005091891158372164, + 0.004656326957046986, + 0.005989740137010813, + 0.009153283201158047, + 0.010083461180329323, + 0.002549855038523674, + 0.003909426275640726, + 0.004234522581100464, + 0.0034345544409006834, + 0.003007977269589901, + 0.0031070879194885492, + 0.009854289703071117, + 0.016419803723692894, + 0.0018837798852473497, + 0.00420552771538496, + 0.007666115649044514, + 0.002487484598532319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low-angle, epic wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly ascend to reveal a vast curtain of water falling over jagged rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rock faces visible as dark, slick with green moss and streaked with dampness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water depicted as a chaotic mix of explosive white foam and translucent streams plunging with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a cinematic style featuring desaturated colors, a fine film grain texture, and an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17234978079795837, + 0.13240206241607666, + 0.19896017014980316, + 0.1319996863603592, + 0.12054535001516342, + 0.0706959143280983, + 0.060270778834819794, + 0.037318769842386246, + 0.04645531252026558, + 0.02900218404829502 + ], + "credit_map": [ + 0.023174269124865532, + 0.022982019931077957, + 0.03410603851079941, + 0.024243902415037155, + 0.016347510740160942, + 0.014633648097515106, + 0.020456096157431602, + 0.016406308859586716, + 0.01784358359873295, + 0.01837415248155594, + 0.025872085243463516, + 0.022482747212052345, + 0.010510014370083809, + 0.011598027311265469, + 0.010629328899085522, + 0.01509213075041771, + 0.020096708089113235, + 0.012312527745962143, + 0.0745815560221672, + 0.06564012169837952, + 0.0060609751380980015, + 0.006083181127905846, + 0.010078608058393002, + 0.004106491804122925, + 0.020098915323615074, + 0.0237693153321743, + 0.017660489305853844, + 0.01931898295879364, + 0.014917110092937946, + 0.010543139651417732, + 0.015457644127309322, + 0.010234097950160503, + 0.01442219503223896, + 0.015445130877196789, + 0.03592263162136078, + 0.02748931758105755, + 0.007465018425136805, + 0.006720616482198238, + 0.008681666105985641, + 0.00439878087490797, + 0.013269961811602116, + 0.01714862696826458, + 0.00891968235373497, + 0.009653209708631039, + 0.005482141859829426, + 0.004875642247498035, + 0.005880212411284447, + 0.005466441623866558, + 0.007380746770650148, + 0.012190783396363258, + 0.01213359646499157, + 0.006388112436980009, + 0.0053918673656880856, + 0.005430309567600489, + 0.006145018618553877, + 0.0052103521302342415, + 0.005753031931817532, + 0.007321426644921303, + 0.004235266707837582, + 0.005294139962643385, + 0.00420069694519043, + 0.002602642634883523, + 0.004872126039117575, + 0.003039439907297492, + 0.0052094366401433945, + 0.006356852129101753, + 0.008894570171833038, + 0.006103350315243006, + 0.0033296928741037846, + 0.006517487578094006, + 0.004051659721881151, + 0.005992268677800894, + 0.00322287087328732, + 0.003580168355256319, + 0.003866655519232154, + 0.005849036853760481, + 0.0024254063609987497, + 0.0030172159895300865, + 0.004846659023314714, + 0.002194174099713564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014191891998052597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_tensor([2473, 2473])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_tensor([2473, 2473])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..da424dfbe07e798997fabe2301a9ad44bcfa7c8c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_tensor([2473, 2473])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, solemn panorama zooms in, plunging us into the depths of a magnificent Baroque library. The space is bathed in a deep blue hue, creating a tranquil and mysterious atmosphere. The camera slowly advances, passing through rows of dark walnut bookshelves reaching up to the soaring vaulted ceiling. These shelves are filled with ancient, leather-bound volumes, their golden spines barely visible in the gloom. As the camera deepens, a vast arched stained-glass window gradually emerges in the center of the frame. A bright, cool cosmic glow filters through, framing an ancient Galilean telescope standing before it as a sharp, backlit silhouette. Its brass tube and intricately carved wooden tripod are reduced to mere outlines in the harsh light, imbued with the solemnity of classical science. At this moment, on the left side of the picture, countless tiny, twinkling starlight particles quietly emerged from the deep shadows. Like a river of glowing cosmic dust, they began to flow slowly in an elegant and mysterious spiral trajectory, carefully condensing towards an invisible center point. The light spots gradually became denser and the light became brighter, as if some unknown form or cosmic mystery was being nurtured in the deep blue tranquility.", + "seed": 5917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5917_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03582347658229992, + "vlm_logprob": -0.03582347658229992, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.03582347658229992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15314632654190063 + ], + "opposite_logprob_tokens": [ + -2.028146266937256 + ], + "logprob_reward": -0.15314632654190063, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15314632654190063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16039811074733734, + 0.09590832144021988, + 0.16952857375144958, + 0.224681556224823, + 0.0653146356344223, + 0.06069029122591019, + 0.09936176985502243, + 0.05697084590792656, + 0.03908321261405945, + 0.028062690049409866 + ], + "credit_map": [ + 0.02119676023721695, + 0.016422973945736885, + 0.032881755381822586, + 0.011770820245146751, + 0.014620383270084858, + 0.021686622872948647, + 0.017307573929429054, + 0.024511223658919334, + 0.009499676525592804, + 0.015977341681718826, + 0.016963670030236244, + 0.00748160295188427, + 0.00927491020411253, + 0.01305303256958723, + 0.008850082755088806, + 0.014808000065386295, + 0.02729140780866146, + 0.024355677887797356, + 0.033773019909858704, + 0.009835605509579182, + 0.011432993225753307, + 0.0277705080807209, + 0.013541266322135925, + 0.021528102457523346, + 0.021976914256811142, + 0.015677744522690773, + 0.06003725901246071, + 0.010908721946179867, + 0.01352067943662405, + 0.06748292595148087, + 0.011965163983404636, + 0.023112140595912933, + 0.011628515087068081, + 0.008191888220608234, + 0.008551724255084991, + 0.004869389347732067, + 0.0068626669235527515, + 0.012897355481982231, + 0.004778585396707058, + 0.007534510921686888, + 0.009333050809800625, + 0.008905510418117046, + 0.009643828496336937, + 0.0051705483347177505, + 0.006427696440368891, + 0.008871182799339294, + 0.004056353587657213, + 0.008282122202217579, + 0.009270566515624523, + 0.02023155800998211, + 0.01615511253476143, + 0.007728708442300558, + 0.009514194913208485, + 0.01899177022278309, + 0.004716988652944565, + 0.012752866372466087, + 0.0069090137258172035, + 0.006001045927405357, + 0.011038282886147499, + 0.0063011664897203445, + 0.0045011043548583984, + 0.007407784461975098, + 0.0039291721768677235, + 0.010883273556828499, + 0.003999688196927309, + 0.004865996073931456, + 0.005269698332995176, + 0.003629949176684022, + 0.004880236927419901, + 0.004891868215054274, + 0.006302242167294025, + 0.005243529565632343, + 0.003958842251449823, + 0.00365143408998847, + 0.00556514598429203, + 0.0023171331267803907, + 0.001820109086111188, + 0.0033482376020401716, + 0.003453347133472562, + 0.0039484393782913685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.15314632654190063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006565084331668913 + ], + "opposite_logprob_tokens": [ + -7.375656604766846 + ], + "logprob_reward": -0.0006565084331668913, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006565084331668913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09773097932338715, + 0.06272630393505096, + 0.14478205144405365, + 0.29812589287757874, + 0.037863727658987045, + 0.03804197162389755, + 0.17927296459674835, + 0.05791325867176056, + 0.04540388286113739, + 0.038138870149850845 + ], + "credit_map": [ + 0.012967740185558796, + 0.00998139102011919, + 0.019804678857326508, + 0.008241514675319195, + 0.010681066662073135, + 0.01315027941018343, + 0.011049771681427956, + 0.011854554526507854, + 0.0097203915938735, + 0.008908959105610847, + 0.013320282101631165, + 0.005149412900209427, + 0.005463486071676016, + 0.008255711756646633, + 0.006718533579260111, + 0.00518953800201416, + 0.02208208665251732, + 0.01778848096728325, + 0.018704727292060852, + 0.00965873058885336, + 0.022002942860126495, + 0.020653175190091133, + 0.01972832903265953, + 0.014163611456751823, + 0.02599896304309368, + 0.009291112422943115, + 0.10129781067371368, + 0.009853706695139408, + 0.01068138051778078, + 0.1099979430437088, + 0.010211649350821972, + 0.02079334668815136, + 0.004655490163713694, + 0.005227915942668915, + 0.004511131439357996, + 0.005227080546319485, + 0.004849639255553484, + 0.004668714012950659, + 0.003504411084577441, + 0.00521934824064374, + 0.00509069487452507, + 0.0068638925440609455, + 0.005095207132399082, + 0.0035297339782118797, + 0.004098759964108467, + 0.005224240478128195, + 0.003363656811416149, + 0.00477579003199935, + 0.017216986045241356, + 0.034506265074014664, + 0.020428970456123352, + 0.014449454843997955, + 0.01877213828265667, + 0.03016217239201069, + 0.009602656587958336, + 0.03413435071706772, + 0.007001645863056183, + 0.008778372779488564, + 0.011166917160153389, + 0.005202161613851786, + 0.004590293392539024, + 0.00738965068012476, + 0.005614493973553181, + 0.008169728331267834, + 0.005883855279535055, + 0.004361965227872133, + 0.007683845702558756, + 0.0036703869700431824, + 0.004108864348381758, + 0.005777366459369659, + 0.006864945869892836, + 0.007052659057080746, + 0.0037912277039140463, + 0.0021536110434681177, + 0.009790285490453243, + 0.0019942810758948326, + 0.0016256084199994802, + 0.005767837632447481, + 0.002223740564659238, + 0.010792279615998268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.0006565084331668913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21987566351890564, + 0.049532219767570496, + 0.19035980105400085, + 0.19399258494377136, + 0.04163846746087074, + 0.04808155819773674, + 0.07309484481811523, + 0.12883630394935608, + 0.03688696026802063, + 0.017701612785458565 + ], + "credit_map": [ + 0.018828626722097397, + 0.022055622190237045, + 0.024597348645329475, + 0.02561955899000168, + 0.02535407617688179, + 0.027106095105409622, + 0.04640967398881912, + 0.02990463376045227, + 0.007491406984627247, + 0.007550318259745836, + 0.006146769504994154, + 0.004839349538087845, + 0.0049736229702830315, + 0.006818945519626141, + 0.00562996556982398, + 0.006081833969801664, + 0.02199118584394455, + 0.021290825679898262, + 0.024065207690000534, + 0.014986167661845684, + 0.029972217977046967, + 0.021960726007819176, + 0.03785283491015434, + 0.018240634351968765, + 0.017173225060105324, + 0.009177014231681824, + 0.062149327248334885, + 0.008119135163724422, + 0.007752120960503817, + 0.06978978216648102, + 0.007462871726602316, + 0.012369079515337944, + 0.004102661274373531, + 0.0038547914009541273, + 0.004938795696943998, + 0.0029791288543492556, + 0.0051854983903467655, + 0.006036372855305672, + 0.004289981443434954, + 0.010251233354210854, + 0.005028969142585993, + 0.006134881637990475, + 0.005642315838485956, + 0.005727386102080345, + 0.006299889180809259, + 0.005155025515705347, + 0.008666513487696648, + 0.005426573567092419, + 0.006448264233767986, + 0.007051898632198572, + 0.016966380178928375, + 0.006998790428042412, + 0.007448859512805939, + 0.00669906148687005, + 0.011355829425156116, + 0.010125752538442612, + 0.011480645276606083, + 0.0102909617125988, + 0.014711243100464344, + 0.014988446608185768, + 0.01343484502285719, + 0.0143672414124012, + 0.030173061415553093, + 0.01938984915614128, + 0.0035752453841269016, + 0.0036510552745312452, + 0.0040194508619606495, + 0.0036362309474498034, + 0.004136356525123119, + 0.004841833841055632, + 0.008309335447847843, + 0.004717450123280287, + 0.0023142171557992697, + 0.0009639805066399276, + 0.005206805653870106, + 0.0009733756305649877, + 0.0007229509647004306, + 0.0017185983015224338, + 0.0008286654483526945, + 0.004973018076270819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01419165637344122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013377059949561954 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013377059949561954, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013377059949561954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20275595784187317, + 0.08036278933286667, + 0.16915808618068695, + 0.1477370262145996, + 0.0598808228969574, + 0.05979039520025253, + 0.09897209703922272, + 0.0532328262925148, + 0.06783513724803925, + 0.06027490645647049 + ], + "credit_map": [ + 0.03588845208287239, + 0.022069796919822693, + 0.0341329462826252, + 0.01668361946940422, + 0.02003927156329155, + 0.03022306226193905, + 0.018327277153730392, + 0.025391530245542526, + 0.009435509331524372, + 0.00976918637752533, + 0.015627525746822357, + 0.007584488019347191, + 0.007188513875007629, + 0.010489093139767647, + 0.01065149251371622, + 0.009616977535188198, + 0.024688927456736565, + 0.02228599414229393, + 0.029002781957387924, + 0.01209692656993866, + 0.015217692591249943, + 0.02693502977490425, + 0.016565963625907898, + 0.022364763543009758, + 0.013815762475132942, + 0.011480833403766155, + 0.03013628162443638, + 0.010319750756025314, + 0.013791023753583431, + 0.03396137058734894, + 0.015253464691340923, + 0.01897851936519146, + 0.009849203750491142, + 0.006303723435848951, + 0.008295021951198578, + 0.007502371445298195, + 0.00743449991568923, + 0.009892795234918594, + 0.004970430862158537, + 0.005632772110402584, + 0.009542129933834076, + 0.0078007737174630165, + 0.010880108922719955, + 0.005195315461605787, + 0.0052093989215791225, + 0.009074393659830093, + 0.004301824141293764, + 0.007786446250975132, + 0.008722418919205666, + 0.010547168552875519, + 0.02889420837163925, + 0.007622339762747288, + 0.00740476930513978, + 0.011842959560453892, + 0.009238574653863907, + 0.01469965185970068, + 0.004735939670354128, + 0.0048007662408053875, + 0.007298734039068222, + 0.005620199721306562, + 0.004967111628502607, + 0.00849124789237976, + 0.00586045254021883, + 0.011458372697234154, + 0.008220940828323364, + 0.007875625044107437, + 0.008674724958837032, + 0.005383341107517481, + 0.005045091267675161, + 0.010491067543625832, + 0.006195973604917526, + 0.01594836637377739, + 0.007242041174322367, + 0.003959889058023691, + 0.01129763200879097, + 0.003964416682720184, + 0.00434090243652463, + 0.008155002258718014, + 0.008056165650486946, + 0.0132588567212224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0013377059949561954 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785185568034649 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009785185568034649, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785185568034649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15639519691467285, + 0.05276850610971451, + 0.10180658847093582, + 0.1167108416557312, + 0.06866561621427536, + 0.04539577290415764, + 0.10147175192832947, + 0.0510932020843029, + 0.16848650574684143, + 0.13720600306987762 + ], + "credit_map": [ + 0.017999136820435524, + 0.018272841349244118, + 0.03349500522017479, + 0.015624532476067543, + 0.015246712602674961, + 0.020473772659897804, + 0.018445026129484177, + 0.01683817245066166, + 0.007339120842516422, + 0.0070251706056296825, + 0.012090673670172691, + 0.0037817922420799732, + 0.0048969099298119545, + 0.00684798089787364, + 0.0057410285808146, + 0.0050458298064768314, + 0.01933361403644085, + 0.014295563101768494, + 0.016847945749759674, + 0.006922798231244087, + 0.010214674286544323, + 0.014158598147332668, + 0.009151984006166458, + 0.010881406255066395, + 0.010679004713892937, + 0.00481854984536767, + 0.03995941951870918, + 0.004437657538801432, + 0.0042312052100896835, + 0.03956378623843193, + 0.004754494875669479, + 0.008266723714768887, + 0.007727312855422497, + 0.005125333089381456, + 0.008644158020615578, + 0.006128670182079077, + 0.007145591080188751, + 0.013059117831289768, + 0.0048084263689816, + 0.016027003526687622, + 0.006148010957986116, + 0.006465784274041653, + 0.0076624611392617226, + 0.0045944503508508205, + 0.004630337003618479, + 0.007193000055849552, + 0.0030096701811999083, + 0.005692057777196169, + 0.009329614229500294, + 0.010706918314099312, + 0.017814958468079567, + 0.009851062670350075, + 0.010480749420821667, + 0.010679686442017555, + 0.01749533787369728, + 0.015113421715795994, + 0.005708134733140469, + 0.004846814088523388, + 0.008164611645042896, + 0.003968044184148312, + 0.00553684588521719, + 0.00793050043284893, + 0.004477533046156168, + 0.010460718534886837, + 0.02426253631711006, + 0.02022709883749485, + 0.022319437935948372, + 0.012927135452628136, + 0.018521562218666077, + 0.01999821327626705, + 0.0246762465685606, + 0.02555427886545658, + 0.016149872913956642, + 0.012106339447200298, + 0.017907744273543358, + 0.011226433329284191, + 0.012878607027232647, + 0.0232707392424345, + 0.023868141695857048, + 0.019798116758465767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.009785185568034649 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn panorama zooms in, plunging us into the depths of a magnificent Baroque library. The space is bathed in a deep blue hue, creating a tranquil and mysterious atmosphere. The camera slowly advances, passing through rows of dark walnut bookshelves reaching up to the soaring vaulted ceiling. These shelves are filled with ancient, leather-bound volumes, their golden spines barely visible in the gloom. As the camera deepens, a vast arched stained-glass window gradually emerges in the center of the frame. A bright, cool cosmic glow filters through, framing an ancient Galilean telescope standing before it as a sharp, backlit silhouette. Its brass tube and intricately carved wooden tripod are reduced to mere outlines in the harsh light, imbued with the solemnity of classical science. At this moment, on the left side of the picture, countless tiny, twinkling starlight particles quietly emerged from the deep shadows. Like a river of glowing cosmic dust, they began to flow slowly in an elegant and mysterious spiral trajectory, carefully condensing towards an invisible center point. The light spots gradually became denser and the light became brighter, as if some unknown form or cosmic mystery was being nurtured in the deep blue tranquility.", + "seed": 5918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5918_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013614439475350081, + "vlm_logprob": -0.013614439475350081, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.013614439475350081, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04168355464935303 + ], + "opposite_logprob_tokens": [ + -3.2916836738586426 + ], + "logprob_reward": -0.04168355464935303, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04168355464935303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32547464966773987, + 0.0537223145365715, + 0.09050586819648743, + 0.10017402470111847, + 0.14356952905654907, + 0.09816545993089676, + 0.06878837198019028, + 0.052845221012830734, + 0.05201548710465431, + 0.01473893504589796 + ], + "credit_map": [ + 0.02061782218515873, + 0.0830894187092781, + 0.0274847112596035, + 0.02002573385834694, + 0.10666176676750183, + 0.023656576871871948, + 0.022161807864904404, + 0.021776841953396797, + 0.00513071846216917, + 0.006703732535243034, + 0.010059107095003128, + 0.005965257994830608, + 0.007236669305711985, + 0.007522607687860727, + 0.005826006177812815, + 0.005278219003230333, + 0.010693001560866833, + 0.011536821722984314, + 0.010770396329462528, + 0.0057775601744651794, + 0.018654176965355873, + 0.014039019122719765, + 0.006086329929530621, + 0.012948575429618359, + 0.008312225341796875, + 0.015070910565555096, + 0.01854691281914711, + 0.010338052175939083, + 0.011204902082681656, + 0.014328054152429104, + 0.007300966884940863, + 0.015072012320160866, + 0.014906712807714939, + 0.009287583641707897, + 0.016664525493979454, + 0.006137842778116465, + 0.014592187479138374, + 0.024038134142756462, + 0.007014589849859476, + 0.05092797428369522, + 0.011232499964535236, + 0.007978530600667, + 0.010092145763337612, + 0.008393103256821632, + 0.011517894454300404, + 0.013773764483630657, + 0.008836084976792336, + 0.02634144201874733, + 0.007856969721615314, + 0.00733646284788847, + 0.010910707525908947, + 0.006750715430825949, + 0.010164064355194569, + 0.010084716603159904, + 0.005470122676342726, + 0.01021462120115757, + 0.006819982081651688, + 0.006322183180600405, + 0.005914722103625536, + 0.0031577509362250566, + 0.00579895218834281, + 0.011257550679147243, + 0.003950444515794516, + 0.009623637422919273, + 0.007317658979445696, + 0.010053696110844612, + 0.004347829148173332, + 0.0036376514472067356, + 0.006838030647486448, + 0.008952208794653416, + 0.0032734691631048918, + 0.007594948634505272, + 0.002023621927946806, + 0.0016954911407083273, + 0.0022729705087840557, + 0.0013427826343104243, + 0.0016642186092212796, + 0.0018853651126846671, + 0.0018168172100558877, + 0.0020376702304929495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.04168355464935303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000647692708298564 + ], + "opposite_logprob_tokens": [ + -7.37564754486084 + ], + "logprob_reward": -0.000647692708298564, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.000647692708298564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09336438775062561, + 0.10599385201931, + 0.0998084545135498, + 0.1254335343837738, + 0.10933617502450943, + 0.21651127934455872, + 0.04939698427915573, + 0.10013303905725479, + 0.04609035328030586, + 0.053931865841150284 + ], + "credit_map": [ + 0.00898727122694254, + 0.011739750392735004, + 0.010207203216850758, + 0.012319337576627731, + 0.01696591265499592, + 0.00733293266966939, + 0.016861407086253166, + 0.008950581774115562, + 0.01183002907782793, + 0.01388939842581749, + 0.016132913529872894, + 0.00844811461865902, + 0.019678644835948944, + 0.016211364418268204, + 0.009808864444494247, + 0.00999453105032444, + 0.012215608730912209, + 0.01556384563446045, + 0.010407700203359127, + 0.010780949145555496, + 0.015761222690343857, + 0.012754730880260468, + 0.008872761391103268, + 0.013451642356812954, + 0.012185521423816681, + 0.014886974357068539, + 0.017405729740858078, + 0.012015465646982193, + 0.012385213747620583, + 0.026388315483927727, + 0.010610751807689667, + 0.019555581733584404, + 0.011998401023447514, + 0.013529106043279171, + 0.014022856950759888, + 0.012390646152198315, + 0.016688013449311256, + 0.013117345981299877, + 0.010917493142187595, + 0.016672329977154732, + 0.025103859603405, + 0.015195761807262897, + 0.03357105329632759, + 0.013012945652008057, + 0.014768281020224094, + 0.03290632367134094, + 0.01210097223520279, + 0.06985209137201309, + 0.006205311510711908, + 0.004948064219206572, + 0.005809841677546501, + 0.007156305946409702, + 0.005570322275161743, + 0.006538869813084602, + 0.005783712957054377, + 0.007384561002254486, + 0.0119550796225667, + 0.008674312382936478, + 0.016242636367678642, + 0.01063647959381342, + 0.017175249755382538, + 0.012595747597515583, + 0.009696713648736477, + 0.013156825676560402, + 0.006554870866239071, + 0.0050428868271410465, + 0.005487516522407532, + 0.0048002880066633224, + 0.006173078902065754, + 0.006620579864829779, + 0.005175529047846794, + 0.006235608831048012, + 0.005676655098795891, + 0.004939639009535313, + 0.00809219665825367, + 0.005506918299943209, + 0.005130045115947723, + 0.008806993253529072, + 0.007574923802167177, + 0.008204500190913677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.000647692708298564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061563044786453 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061563044786453, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061563044786453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07493274658918381, + 0.15711578726768494, + 0.02541176602244377, + 0.1272927075624466, + 0.19817787408828735, + 0.14559657871723175, + 0.15139168500900269, + 0.035925302654504776, + 0.05667389556765556, + 0.027481624856591225 + ], + "credit_map": [ + 0.004720586817711592, + 0.017523009330034256, + 0.0050298795104026794, + 0.007815673016011715, + 0.02135039120912552, + 0.004683052655309439, + 0.008286349475383759, + 0.005523794796317816, + 0.022217126563191414, + 0.016437187790870667, + 0.020865583792328835, + 0.017483869567513466, + 0.02526729181408882, + 0.024398047477006912, + 0.017394738271832466, + 0.013051948510110378, + 0.0027078839484602213, + 0.0034286994487047195, + 0.003128663171082735, + 0.0030095481779426336, + 0.003722970373928547, + 0.002753553679212928, + 0.003034282010048628, + 0.003626164747402072, + 0.013942888006567955, + 0.012530094012618065, + 0.015686463564634323, + 0.00955966953188181, + 0.012194119393825531, + 0.03102017194032669, + 0.010186144150793552, + 0.02217315509915352, + 0.019386637955904007, + 0.012834064662456512, + 0.021240491420030594, + 0.008874699473381042, + 0.021965060383081436, + 0.033202096819877625, + 0.010830715298652649, + 0.06984411180019379, + 0.016497913748025894, + 0.012658028863370419, + 0.018162066116929054, + 0.009088539518415928, + 0.01567731983959675, + 0.02105836011469364, + 0.010235880501568317, + 0.042218469083309174, + 0.015484411269426346, + 0.014970810152590275, + 0.01676495186984539, + 0.01984473504126072, + 0.01967773213982582, + 0.019894972443580627, + 0.02307969518005848, + 0.02167437970638275, + 0.008016201667487621, + 0.0037026770878583193, + 0.004377803299576044, + 0.002307446440681815, + 0.004484033677726984, + 0.004848394077271223, + 0.002602925291284919, + 0.005585821345448494, + 0.007101759314537048, + 0.004203971941024065, + 0.005723604001104832, + 0.004507226403802633, + 0.005002276506274939, + 0.013759283348917961, + 0.005359814502298832, + 0.011015959084033966, + 0.0038499166257679462, + 0.002066168235614896, + 0.0038487212732434273, + 0.0023754537105560303, + 0.002665136707946658, + 0.004437262192368507, + 0.003560159355401993, + 0.004678808152675629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.016061563044786453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010413468116894364 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010413468116894364, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010413468116894364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23378494381904602, + 0.06781724840402603, + 0.059755921363830566, + 0.18831582367420197, + 0.12550313770771027, + 0.03525358438491821, + 0.1070411205291748, + 0.06515531241893768, + 0.06301920115947723, + 0.054353728890419006 + ], + "credit_map": [ + 0.0147785022854805, + 0.061025239527225494, + 0.01681034453213215, + 0.015713918954133987, + 0.07754715532064438, + 0.016334762796759605, + 0.016970057040452957, + 0.014604945667088032, + 0.0072434162721037865, + 0.007998274639248848, + 0.009078356437385082, + 0.007951891981065273, + 0.012936166487634182, + 0.008913077414035797, + 0.007913593202829361, + 0.005782464984804392, + 0.0056547001004219055, + 0.010028649121522903, + 0.006260020192712545, + 0.0059067667461931705, + 0.01339565310627222, + 0.006156305782496929, + 0.005422353744506836, + 0.0069314646534621716, + 0.019742753356695175, + 0.020602963864803314, + 0.02572789043188095, + 0.01309273112565279, + 0.019220247864723206, + 0.048987437039613724, + 0.013239442370831966, + 0.027702348306775093, + 0.012974015437066555, + 0.0072886585257947445, + 0.01520618237555027, + 0.0056691356003284454, + 0.008993670344352722, + 0.022176247090101242, + 0.0064104292541742325, + 0.04678479582071304, + 0.0027160572353750467, + 0.0050955419428646564, + 0.003847169689834118, + 0.003400997957214713, + 0.0066447495482862, + 0.004391499795019627, + 0.0031749021727591753, + 0.005982663482427597, + 0.010083535686135292, + 0.009096130728721619, + 0.016272034496068954, + 0.00983815360814333, + 0.0147390216588974, + 0.021952180191874504, + 0.009498088620603085, + 0.015561969950795174, + 0.011986011639237404, + 0.0074944449588656425, + 0.005636345129460096, + 0.0040923538617789745, + 0.011497952975332737, + 0.01056703645735979, + 0.004785323515534401, + 0.00909583643078804, + 0.013986882753670216, + 0.005500349681824446, + 0.007711499929428101, + 0.0034317944664508104, + 0.00520899286493659, + 0.012418040074408054, + 0.004205681383609772, + 0.010555950924754143, + 0.005731991492211819, + 0.006472162902355194, + 0.0060862405225634575, + 0.004524842370301485, + 0.005371997598558664, + 0.009900929406285286, + 0.0066320644691586494, + 0.009633494541049004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0010413468116894364 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008638040162622929, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638040162622929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10312222689390182, + 0.043587826192379, + 0.0501786507666111, + 0.09527641534805298, + 0.03613778203725815, + 0.1285337507724762, + 0.3466934263706207, + 0.0370425321161747, + 0.03314175456762314, + 0.126285582780838 + ], + "credit_map": [ + 0.006406488828361034, + 0.02624458819627762, + 0.006540399976074696, + 0.007254220079630613, + 0.03379549831151962, + 0.0076371910981833935, + 0.008668449707329273, + 0.006575396284461021, + 0.004315290600061417, + 0.005742070730775595, + 0.0055144354701042175, + 0.00442891800776124, + 0.008721436373889446, + 0.005854369606822729, + 0.0051602451130747795, + 0.0038510619197040796, + 0.005360961426049471, + 0.008171153254806995, + 0.006357569247484207, + 0.004339313600212336, + 0.010332928039133549, + 0.0059163388796150684, + 0.0032132857013493776, + 0.006487100385129452, + 0.009739872999489307, + 0.011289721354842186, + 0.011823075823485851, + 0.007184573449194431, + 0.009156324900686741, + 0.024510901421308517, + 0.008041813969612122, + 0.013530128635466099, + 0.004275269340723753, + 0.004237314686179161, + 0.008046370930969715, + 0.0027076860424131155, + 0.004182832315564156, + 0.004584015347063541, + 0.0019792986568063498, + 0.006124993320554495, + 0.012164890766143799, + 0.014561324380338192, + 0.01224627997726202, + 0.009468451142311096, + 0.021030165255069733, + 0.01775718852877617, + 0.009564923122525215, + 0.03174052760004997, + 0.03636613488197327, + 0.028109002858400345, + 0.060798950493335724, + 0.03473035246133804, + 0.03957219049334526, + 0.0646781325340271, + 0.02931985631585121, + 0.05311879888176918, + 0.007906808517873287, + 0.003434182843193412, + 0.003005298553034663, + 0.0029139628168195486, + 0.004988839849829674, + 0.006236364599317312, + 0.003300267504528165, + 0.005256802774965763, + 0.006072251591831446, + 0.0032386090606451035, + 0.003746413392946124, + 0.001841518678702414, + 0.0029260481242090464, + 0.007617764640599489, + 0.0020379626657813787, + 0.005661188159137964, + 0.012015461921691895, + 0.011083131656050682, + 0.012781709432601929, + 0.009617960080504417, + 0.01082870177924633, + 0.0287633016705513, + 0.014758800156414509, + 0.026436522603034973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.008638040162622929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn panorama zooms in, plunging us into the depths of a magnificent Baroque library. The space is bathed in a deep blue hue, creating a tranquil and mysterious atmosphere. The camera slowly advances, passing through rows of dark walnut bookshelves reaching up to the soaring vaulted ceiling. These shelves are filled with ancient, leather-bound volumes, their golden spines barely visible in the gloom. As the camera deepens, a vast arched stained-glass window gradually emerges in the center of the frame. A bright, cool cosmic glow filters through, framing an ancient Galilean telescope standing before it as a sharp, backlit silhouette. Its brass tube and intricately carved wooden tripod are reduced to mere outlines in the harsh light, imbued with the solemnity of classical science. At this moment, on the left side of the picture, countless tiny, twinkling starlight particles quietly emerged from the deep shadows. Like a river of glowing cosmic dust, they began to flow slowly in an elegant and mysterious spiral trajectory, carefully condensing towards an invisible center point. The light spots gradually became denser and the light became brighter, as if some unknown form or cosmic mystery was being nurtured in the deep blue tranquility.", + "seed": 5919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5919_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05406622774316929, + "vlm_logprob": -0.05406622774316929, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.05406622774316929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2561618387699127 + ], + "opposite_logprob_tokens": [ + -1.5061618089675903 + ], + "logprob_reward": -0.2561618387699127, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2561618387699127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09253406524658203, + 0.046161752194166183, + 0.6261717677116394, + 0.06764237582683563, + 0.03675828129053116, + 0.023957233875989914, + 0.01623421721160412, + 0.02239050716161728, + 0.056160517036914825, + 0.011989288963377476 + ], + "credit_map": [ + 0.013028050772845745, + 0.010843273252248764, + 0.015896108001470566, + 0.01598306931555271, + 0.01319377962499857, + 0.008815446868538857, + 0.007489799987524748, + 0.007284531369805336, + 0.006937343627214432, + 0.006787551566958427, + 0.005900419782847166, + 0.0048567247577011585, + 0.006353768985718489, + 0.005678755231201649, + 0.0028331452049314976, + 0.0068140472285449505, + 0.13827592134475708, + 0.07171820849180222, + 0.045052967965602875, + 0.015183674171566963, + 0.03655684366822243, + 0.2392825335264206, + 0.0171065516769886, + 0.06299509108066559, + 0.008094608783721924, + 0.008298012427985668, + 0.00937199592590332, + 0.004801595583558083, + 0.008193565532565117, + 0.012269413098692894, + 0.0043800282292068005, + 0.01223315391689539, + 0.008688422851264477, + 0.003755957121029496, + 0.00481484504416585, + 0.001954505918547511, + 0.002665630541741848, + 0.007497122976928949, + 0.002166654448956251, + 0.005215141456574202, + 0.0037587943952530622, + 0.0021981883328408003, + 0.004553563892841339, + 0.0018572505796328187, + 0.0022879887837916613, + 0.0039816610515117645, + 0.0017663374310359359, + 0.0035534463822841644, + 0.0017110728658735752, + 0.0008190816733986139, + 0.002145073376595974, + 0.0009653748129494488, + 0.001216679229401052, + 0.005189664661884308, + 0.0009798795217648149, + 0.0032073911279439926, + 0.0031018946319818497, + 0.0038119887467473745, + 0.0025570555590093136, + 0.003118356456980109, + 0.0024068194907158613, + 0.002752841217443347, + 0.002220414811745286, + 0.002421134850010276, + 0.005460426677018404, + 0.003649947000667453, + 0.008461924269795418, + 0.011213327758014202, + 0.004210094455629587, + 0.007339250296354294, + 0.00633544847369194, + 0.00949009694159031, + 0.0017882516840472817, + 0.001282513840124011, + 0.0016783862374722958, + 0.0018933353712782264, + 0.0010474794544279575, + 0.0015827586175873876, + 0.0014530239859595895, + 0.0012635407038033009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.2561618387699127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00039939055568538606 + ], + "opposite_logprob_tokens": [ + -7.875399589538574 + ], + "logprob_reward": -0.00039939055568538606, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00039939055568538606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06963518261909485, + 0.03548299893736839, + 0.41515976190567017, + 0.055310022085905075, + 0.12722189724445343, + 0.06672195345163345, + 0.10867840051651001, + 0.03389837592840195, + 0.06761454045772552, + 0.0202768687158823 + ], + "credit_map": [ + 0.008398424834012985, + 0.010013408958911896, + 0.010015169158577919, + 0.01092202216386795, + 0.012171671725809574, + 0.006793199107050896, + 0.006452724803239107, + 0.004868559073656797, + 0.004949360620230436, + 0.005135572515428066, + 0.0035881998483091593, + 0.00297339865937829, + 0.004891162272542715, + 0.005189183633774519, + 0.002692839363589883, + 0.006063283421099186, + 0.09007735550403595, + 0.048375051468610764, + 0.02963007055222988, + 0.009899214841425419, + 0.024864234030246735, + 0.15908432006835938, + 0.011318203061819077, + 0.041911330074071884, + 0.006895946338772774, + 0.008087528869509697, + 0.008680337108671665, + 0.004618385341018438, + 0.005858655087649822, + 0.00798184797167778, + 0.004167958162724972, + 0.009019357152283192, + 0.03853633254766464, + 0.009149610064923763, + 0.016505910083651543, + 0.005031779408454895, + 0.005450557451695204, + 0.03508760407567024, + 0.004100278485566378, + 0.013359826058149338, + 0.011042958125472069, + 0.0047140601091086864, + 0.011449849233031273, + 0.005130194593220949, + 0.00752980588003993, + 0.009186760522425175, + 0.0073768445290625095, + 0.010291481390595436, + 0.014377165585756302, + 0.007522539235651493, + 0.015423542819917202, + 0.009025153703987598, + 0.007686375640332699, + 0.030763456597924232, + 0.0070227994583547115, + 0.016857372596859932, + 0.003383658593520522, + 0.005571819841861725, + 0.0033833261113613844, + 0.0065902648493647575, + 0.0030830143950879574, + 0.003592531895264983, + 0.003955455031245947, + 0.004338303115218878, + 0.009916807524859905, + 0.006141044665127993, + 0.016096172854304314, + 0.010112496092915535, + 0.004030926618725061, + 0.0077858613803982735, + 0.007296593859791756, + 0.006234633736312389, + 0.00382899958640337, + 0.001097212079912424, + 0.004472028464078903, + 0.0020459226798266172, + 0.0011799378553405404, + 0.003040195908397436, + 0.0019453235436230898, + 0.002667248249053955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.00039939055568538606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636031299829483 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008636031299829483, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636031299829483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13202817738056183, + 0.07589422911405563, + 0.11353631317615509, + 0.13319672644138336, + 0.14097709953784943, + 0.1753927320241928, + 0.10272887349128723, + 0.027490131556987762, + 0.08482146263122559, + 0.013934269547462463 + ], + "credit_map": [ + 0.017572788521647453, + 0.017878469079732895, + 0.022330526262521744, + 0.010338225401937962, + 0.034324564039707184, + 0.010364240035414696, + 0.010607311502099037, + 0.008612066507339478, + 0.009801363572478294, + 0.008699084632098675, + 0.013577179051935673, + 0.006121112033724785, + 0.010737585835158825, + 0.008849428035318851, + 0.006337854545563459, + 0.011770624667406082, + 0.019065048545598984, + 0.00982986856251955, + 0.015044624917209148, + 0.006588807795196772, + 0.009657176211476326, + 0.03197575733065605, + 0.005639171693474054, + 0.015735862776637077, + 0.013110407628118992, + 0.016033954918384552, + 0.019214386120438576, + 0.008638466708362103, + 0.014956849627196789, + 0.03664804622530937, + 0.009376495145261288, + 0.015218137763440609, + 0.04187113419175148, + 0.010800080373883247, + 0.018319327384233475, + 0.00492778280749917, + 0.0067426953464746475, + 0.036954980343580246, + 0.005606690421700478, + 0.01575443148612976, + 0.03613455966114998, + 0.012432382442057133, + 0.03465784713625908, + 0.011079905554652214, + 0.014319333247840405, + 0.024780893698334694, + 0.012714224867522717, + 0.02927357703447342, + 0.010535622015595436, + 0.006053437478840351, + 0.012678622268140316, + 0.0061780475080013275, + 0.006451848428696394, + 0.032233260571956635, + 0.004758363123983145, + 0.023839667439460754, + 0.0033251852728426456, + 0.0034713675267994404, + 0.003890139516443014, + 0.0034121593926101923, + 0.004747691098600626, + 0.0030441798735409975, + 0.002621884923428297, + 0.002977527678012848, + 0.014336763881146908, + 0.005434668622910976, + 0.023458339273929596, + 0.00966534111648798, + 0.0042867339216172695, + 0.01031963899731636, + 0.008119294419884682, + 0.00920068472623825, + 0.00192590756341815, + 0.0016981350490823388, + 0.0017311156261712313, + 0.002788197249174118, + 0.0013724552700296044, + 0.0019540400244295597, + 0.0010868406388908625, + 0.0013775791740044951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.008636031299829483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010408704401925206 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010408704401925206, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010408704401925206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10146396607160568, + 0.07491595298051834, + 0.4396204948425293, + 0.08063660562038422, + 0.05035412684082985, + 0.08214809000492096, + 0.044682908803224564, + 0.02650676667690277, + 0.06520020216703415, + 0.03447084501385689 + ], + "credit_map": [ + 0.01267945859581232, + 0.016336053609848022, + 0.01409133430570364, + 0.011042008176445961, + 0.021965285763144493, + 0.008407131768763065, + 0.009790456853806973, + 0.00715224351733923, + 0.010540368035435677, + 0.010104229673743248, + 0.007312798406928778, + 0.007188324350863695, + 0.009875291958451271, + 0.010490849614143372, + 0.008583231829106808, + 0.010820860043168068, + 0.09753672778606415, + 0.049316905438899994, + 0.03201925382018089, + 0.010739780031144619, + 0.026631075888872147, + 0.1673012673854828, + 0.012194767594337463, + 0.04388069733977318, + 0.009274568408727646, + 0.009940202347934246, + 0.013056237250566483, + 0.00718696927651763, + 0.010131282731890678, + 0.011003607884049416, + 0.006706136744469404, + 0.013337597250938416, + 0.0097909364849329, + 0.006367980036884546, + 0.009386072866618633, + 0.0035532827023416758, + 0.00308780069462955, + 0.007887255400419235, + 0.00308629940263927, + 0.007194502279162407, + 0.016218174248933792, + 0.0064888764172792435, + 0.014309141784906387, + 0.007627464830875397, + 0.006588401738554239, + 0.011409986764192581, + 0.0066901808604598045, + 0.012815861962735653, + 0.0064314017072319984, + 0.004571770783513784, + 0.007789416704326868, + 0.0040303911082446575, + 0.004371643532067537, + 0.008770819753408432, + 0.00317923235706985, + 0.005538232158869505, + 0.002861643210053444, + 0.004497713875025511, + 0.0025271365884691477, + 0.003419886576011777, + 0.002755821216851473, + 0.003042961237952113, + 0.0027993768453598022, + 0.004602224566042423, + 0.007298513315618038, + 0.007800254039466381, + 0.007630355656147003, + 0.011418611742556095, + 0.005064177326858044, + 0.007985283620655537, + 0.010452008806169033, + 0.007550996262580156, + 0.004526550881564617, + 0.004529313649982214, + 0.004354918375611305, + 0.006936708465218544, + 0.003037945833057165, + 0.004047772381454706, + 0.0032568012829869986, + 0.003780832514166832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0010408704401925206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093007650226355 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093007650226355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093007650226355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08636785298585892, + 0.04367780685424805, + 0.3264051377773285, + 0.02866164594888687, + 0.031354088336229324, + 0.04930371046066284, + 0.051060836762189865, + 0.09055393189191818, + 0.13055752217769623, + 0.16205747425556183 + ], + "credit_map": [ + 0.011428650468587875, + 0.011980275623500347, + 0.014889892190694809, + 0.009679412469267845, + 0.0184236541390419, + 0.005662587471306324, + 0.006766248494386673, + 0.0075371298007667065, + 0.004214102867990732, + 0.005950794089585543, + 0.003887313650920987, + 0.005106760188937187, + 0.006078760605305433, + 0.0057717240415513515, + 0.006030390039086342, + 0.0066379643976688385, + 0.07196547091007233, + 0.036730535328388214, + 0.02341669611632824, + 0.007498805411159992, + 0.019506599754095078, + 0.12500058114528656, + 0.009233077056705952, + 0.03305336460471153, + 0.003636403940618038, + 0.00349545874632895, + 0.004957466386258602, + 0.0024361007381230593, + 0.0033568288199603558, + 0.004263713490217924, + 0.002811363898217678, + 0.00370430969633162, + 0.006495187990367413, + 0.002803756855428219, + 0.0064115626737475395, + 0.0018103292677551508, + 0.0019182009855285287, + 0.006125608459115028, + 0.0018638584297150373, + 0.003925584256649017, + 0.011357656680047512, + 0.0033154794946312904, + 0.009242655709385872, + 0.0036009864415973425, + 0.00409594364464283, + 0.006722916848957539, + 0.00424418319016695, + 0.006723891012370586, + 0.006614205427467823, + 0.0036196294240653515, + 0.0070483083836734295, + 0.004626098554581404, + 0.004908896051347256, + 0.013048043474555016, + 0.003291559172794223, + 0.007904097437858582, + 0.008919840678572655, + 0.013213057070970535, + 0.009355733171105385, + 0.016160471364855766, + 0.009024227038025856, + 0.011159964837133884, + 0.010937306098639965, + 0.011783331632614136, + 0.014787525869905949, + 0.02666621096432209, + 0.012020890600979328, + 0.020857559517025948, + 0.010531622916460037, + 0.014701133593916893, + 0.017588332295417786, + 0.013404246419668198, + 0.022315239533782005, + 0.022316791117191315, + 0.022787943482398987, + 0.0343024842441082, + 0.013417947106063366, + 0.018519269302487373, + 0.013232073746621609, + 0.015165739692747593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.004093007650226355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, solemn panorama zooms in, plunging us into the depths of a magnificent Baroque library. The space is bathed in a deep blue hue, creating a tranquil and mysterious atmosphere. The camera slowly advances, passing through rows of dark walnut bookshelves reaching up to the soaring vaulted ceiling. These shelves are filled with ancient, leather-bound volumes, their golden spines barely visible in the gloom. As the camera deepens, a vast arched stained-glass window gradually emerges in the center of the frame. A bright, cool cosmic glow filters through, framing an ancient Galilean telescope standing before it as a sharp, backlit silhouette. Its brass tube and intricately carved wooden tripod are reduced to mere outlines in the harsh light, imbued with the solemnity of classical science. At this moment, on the left side of the picture, countless tiny, twinkling starlight particles quietly emerged from the deep shadows. Like a river of glowing cosmic dust, they began to flow slowly in an elegant and mysterious spiral trajectory, carefully condensing towards an invisible center point. The light spots gradually became denser and the light became brighter, as if some unknown form or cosmic mystery was being nurtured in the deep blue tranquility.", + "seed": 5920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_6/index_2473_seed_5920_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023140291951131077, + "vlm_logprob": -0.023140291951131077, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.023140291951131077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08731415867805481 + ], + "opposite_logprob_tokens": [ + -2.5873141288757324 + ], + "logprob_reward": -0.08731415867805481, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08731415867805481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2653889060020447, + 0.04007068648934364, + 0.05618104338645935, + 0.1644020676612854, + 0.02654416859149933, + 0.1027837023139, + 0.07981913536787033, + 0.1758164018392563, + 0.05339584872126579, + 0.035598088055849075 + ], + "credit_map": [ + 0.019703222438693047, + 0.021005896851420403, + 0.02448819950222969, + 0.0196601003408432, + 0.015871455892920494, + 0.0967906191945076, + 0.023507582023739815, + 0.044361814856529236, + 0.004321955144405365, + 0.004348165821284056, + 0.003862926736474037, + 0.006709115579724312, + 0.0032155762892216444, + 0.00457091024145484, + 0.004111877176910639, + 0.00893015693873167, + 0.008639383129775524, + 0.004858235362917185, + 0.005683737341314554, + 0.00805183220654726, + 0.0055688596330583096, + 0.00894135795533657, + 0.005688345059752464, + 0.008749284781515598, + 0.04476885497570038, + 0.00949701014906168, + 0.04197860509157181, + 0.008315479382872581, + 0.01081038173288107, + 0.02461227774620056, + 0.007390180137008429, + 0.017029263079166412, + 0.0036628679372370243, + 0.0021607999224215746, + 0.004095642361789942, + 0.0027886941097676754, + 0.003810445312410593, + 0.004281464498490095, + 0.00233569392003119, + 0.0034085563383996487, + 0.010216263122856617, + 0.010326847434043884, + 0.01308421976864338, + 0.016900038346648216, + 0.01502353884279728, + 0.011612290516495705, + 0.010621508583426476, + 0.014998982660472393, + 0.013271043077111244, + 0.006159788463264704, + 0.009293396025896072, + 0.008867134340107441, + 0.007792782038450241, + 0.011736595071852207, + 0.00629069609567523, + 0.016407687216997147, + 0.03691495582461357, + 0.016220642253756523, + 0.04523297771811485, + 0.013121390715241432, + 0.020261213183403015, + 0.016241509467363358, + 0.0128086619079113, + 0.015015033073723316, + 0.006919453386217356, + 0.004249812103807926, + 0.005379895213991404, + 0.003377228509634733, + 0.005879820790141821, + 0.01577521488070488, + 0.004200061317533255, + 0.0076143620535731316, + 0.008055672980844975, + 0.002821286441758275, + 0.00232379836961627, + 0.003174968296661973, + 0.005765244830399752, + 0.00423825066536665, + 0.004035774152725935, + 0.005183087661862373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.08731415867805481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003195490571670234 + ], + "opposite_logprob_tokens": [ + -8.125319480895996 + ], + "logprob_reward": -0.0003195490571670234, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003195490571670234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13445471227169037, + 0.028922466561198235, + 0.05857985466718674, + 0.044832952320575714, + 0.02260979264974594, + 0.017024148255586624, + 0.040757112205028534, + 0.5760644674301147, + 0.030199376866221428, + 0.04655512049794197 + ], + "credit_map": [ + 0.008650144562125206, + 0.01098877377808094, + 0.010603014379739761, + 0.0056477137841284275, + 0.010025289840996265, + 0.05450313538312912, + 0.008624143898487091, + 0.02541249617934227, + 0.003048060229048133, + 0.0030894456431269646, + 0.002397493924945593, + 0.005587542429566383, + 0.0032878615893423557, + 0.003744285088032484, + 0.0035573176573961973, + 0.004210459068417549, + 0.00992746651172638, + 0.005249767564237118, + 0.006576361134648323, + 0.00852134544402361, + 0.005440997891128063, + 0.00749635323882103, + 0.005842363927513361, + 0.00952520128339529, + 0.01092467550188303, + 0.0025924888905137777, + 0.011834464967250824, + 0.0028191690798848867, + 0.0027901651337742805, + 0.00687125138938427, + 0.002262655645608902, + 0.0047380863688886166, + 0.004068463109433651, + 0.0012233121087774634, + 0.006772794760763645, + 0.0019645155407488346, + 0.0015837432583793998, + 0.002576169790700078, + 0.0013701742282137275, + 0.0030506199691444635, + 0.0017725940560922027, + 0.0013554197503253818, + 0.002066093496978283, + 0.003407241078093648, + 0.0016325439792126417, + 0.0020568736363202333, + 0.0020132448989897966, + 0.0027201364282518625, + 0.009110521525144577, + 0.002873106161132455, + 0.005750703625380993, + 0.003646508790552616, + 0.00424025347456336, + 0.005772992968559265, + 0.002807667013257742, + 0.006555357947945595, + 0.13060534000396729, + 0.04663366079330444, + 0.15794409811496735, + 0.03730368986725807, + 0.06816628575325012, + 0.05568753555417061, + 0.0335666798055172, + 0.046157196164131165, + 0.004687799606472254, + 0.0024793685879558325, + 0.0029978391248732805, + 0.0023556652013212442, + 0.0029021156951785088, + 0.0077968863770365715, + 0.0028148831333965063, + 0.004164819605648518, + 0.008233520202338696, + 0.003790680319070816, + 0.0048173461109399796, + 0.006474884692579508, + 0.005242838058620691, + 0.004817526321858168, + 0.005427827127277851, + 0.0077504962682724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0003195490571670234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536060065031052 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012536060065031052, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536060065031052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1347794085741043, + 0.11370127648115158, + 0.12560318410396576, + 0.09279697388410568, + 0.1297529935836792, + 0.08712015300989151, + 0.10761917382478714, + 0.08470111340284348, + 0.029100533574819565, + 0.09482520073652267 + ], + "credit_map": [ + 0.009102742187678814, + 0.015986168757081032, + 0.009766586124897003, + 0.023889310657978058, + 0.010510670021176338, + 0.0314144492149353, + 0.011292815208435059, + 0.022816665470600128, + 0.010149665176868439, + 0.010831771418452263, + 0.011956239119172096, + 0.02229355089366436, + 0.010271696373820305, + 0.01652347669005394, + 0.01550935860723257, + 0.016165517270565033, + 0.014120155945420265, + 0.011827075853943825, + 0.012544993311166763, + 0.02446000650525093, + 0.011227406561374664, + 0.015031079761683941, + 0.015618282370269299, + 0.02077419124543667, + 0.023113839328289032, + 0.005883018020540476, + 0.019376544281840324, + 0.008011719211935997, + 0.005593268200755119, + 0.011881380341947079, + 0.007946263067424297, + 0.010990944691002369, + 0.021681608632206917, + 0.01075179036706686, + 0.03382060304284096, + 0.008047540672123432, + 0.010606477037072182, + 0.018368039280176163, + 0.007391728460788727, + 0.019085202366113663, + 0.009485665708780289, + 0.009533299133181572, + 0.011484208516776562, + 0.013850039802491665, + 0.012640426866710186, + 0.010302633978426456, + 0.008500608615577221, + 0.01132327038794756, + 0.014895838685333729, + 0.010188604705035686, + 0.017204072326421738, + 0.016486238688230515, + 0.011141260154545307, + 0.012095722369849682, + 0.012308817356824875, + 0.013298624195158482, + 0.0164524894207716, + 0.006238596513867378, + 0.015728386119008064, + 0.0067795803770422935, + 0.009119933471083641, + 0.010277632623910904, + 0.009595006704330444, + 0.010509485378861427, + 0.0038478521164506674, + 0.002266781870275736, + 0.0036009580362588167, + 0.0029506662394851446, + 0.0025389359798282385, + 0.005475890357047319, + 0.0034817007835954428, + 0.004937748424708843, + 0.019271159544587135, + 0.005327632185071707, + 0.0062725963070988655, + 0.010068870149552822, + 0.012415892444550991, + 0.010909473523497581, + 0.011837020516395569, + 0.01872256211936474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.012536060065031052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08034045994281769, + 0.06799332052469254, + 0.08167994767427444, + 0.14395982027053833, + 0.04231378063559532, + 0.04802098125219345, + 0.11172787845134735, + 0.08565167337656021, + 0.11092397570610046, + 0.2273881435394287 + ], + "credit_map": [ + 0.008048304356634617, + 0.006898422725498676, + 0.013856538571417332, + 0.00859135203063488, + 0.006860572844743729, + 0.0216265507042408, + 0.00764872133731842, + 0.00680999830365181, + 0.005850757472217083, + 0.006774636451154947, + 0.0069238729774951935, + 0.012010430917143822, + 0.006665047258138657, + 0.009096450172364712, + 0.0075237564742565155, + 0.013148369267582893, + 0.007298346608877182, + 0.007550718262791634, + 0.0070450082421302795, + 0.020589936524629593, + 0.005594808608293533, + 0.010743271559476852, + 0.009162185713648796, + 0.013695670291781425, + 0.040254946798086166, + 0.008047451265156269, + 0.037789322435855865, + 0.007281624712049961, + 0.008466722443699837, + 0.020686689764261246, + 0.006368993315845728, + 0.015064057894051075, + 0.005081881303340197, + 0.0029055853374302387, + 0.0076467241160571575, + 0.007097756490111351, + 0.004212219268083572, + 0.0063304840587079525, + 0.004037234932184219, + 0.005001890938729048, + 0.004888258408755064, + 0.004377888049930334, + 0.0057104057632386684, + 0.010043498128652573, + 0.005453862715512514, + 0.00572900241240859, + 0.004531921353191137, + 0.007286141626536846, + 0.02151600643992424, + 0.008615553379058838, + 0.011408681981265545, + 0.012758446857333183, + 0.009857339784502983, + 0.01716121844947338, + 0.007502256892621517, + 0.02290838398039341, + 0.015360581688582897, + 0.007007128559052944, + 0.018121743574738503, + 0.006652669049799442, + 0.009220652282238007, + 0.008466611616313457, + 0.01117674633860588, + 0.009645541198551655, + 0.013707596808671951, + 0.009725715033710003, + 0.010643464513123035, + 0.01506783813238144, + 0.012897025793790817, + 0.01853012479841709, + 0.013523167930543423, + 0.01682904176414013, + 0.04914780706167221, + 0.014132142066955566, + 0.01465522963553667, + 0.02337520197033882, + 0.030798546969890594, + 0.024889785796403885, + 0.03104383498430252, + 0.03934558853507042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.0013360392767935991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a Baroque library with dark walnut bookshelves and a vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the overall lighting bathed in a deep blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly advance (zoom in) from the entrance towards the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast arched stained-glass window visible in the center of the frame with a bright cosmic glow filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is an ancient Galilean telescope visible as a backlit silhouette with a brass tube and wooden tripod?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04940566048026085, + 0.03484788164496422, + 0.025128506124019623, + 0.1366577446460724, + 0.021850446239113808, + 0.01821376569569111, + 0.045603737235069275, + 0.0535384826362133, + 0.04768974334001541, + 0.567064106464386 + ], + "credit_map": [ + 0.0036482850555330515, + 0.004573953803628683, + 0.003596802009269595, + 0.004468981176614761, + 0.00495271198451519, + 0.016391286626458168, + 0.00547671876847744, + 0.006296919658780098, + 0.004477685783058405, + 0.003920062445104122, + 0.006120619829744101, + 0.005091750528663397, + 0.004370515700429678, + 0.003998317755758762, + 0.0032676060218364, + 0.0036013193894177675, + 0.004779904615134001, + 0.0032034090254455805, + 0.0032754684798419476, + 0.0028004369232803583, + 0.0025296914391219616, + 0.0036096135154366493, + 0.0019473478896543384, + 0.002982632489874959, + 0.03793640807271004, + 0.006694953888654709, + 0.034684352576732635, + 0.006305077113211155, + 0.008815686218440533, + 0.021031128242611885, + 0.006436091847717762, + 0.014754034578800201, + 0.003961590584367514, + 0.0014518146635964513, + 0.004640015307813883, + 0.001933434745296836, + 0.0016965443501248956, + 0.004009895958006382, + 0.0012951105600222945, + 0.0028620371595025063, + 0.001691916724666953, + 0.0021017238032072783, + 0.002441737800836563, + 0.0026100552640855312, + 0.0025708016473799944, + 0.002486197976395488, + 0.001992839854210615, + 0.0023184900637716055, + 0.007167408242821693, + 0.0040292805060744286, + 0.005920607130974531, + 0.006299363449215889, + 0.004318589810281992, + 0.006049235351383686, + 0.0039262790232896805, + 0.007892968133091927, + 0.00984334759414196, + 0.0038225073367357254, + 0.012785043567419052, + 0.0032823276706039906, + 0.005796326790004969, + 0.005791554693132639, + 0.0039664446376264095, + 0.008250929415225983, + 0.005928385071456432, + 0.00335553172044456, + 0.0029086207505315542, + 0.0029032777529209852, + 0.005834250245243311, + 0.008580203168094158, + 0.007891122251749039, + 0.010288349352777004, + 0.11768267303705215, + 0.036085452884435654, + 0.038047246634960175, + 0.06219855323433876, + 0.07488732039928436, + 0.06149415299296379, + 0.07881253957748413, + 0.09785609692335129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.014195652678608894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_tensor([9993, 9993])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_tensor([9993, 9993])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8181baf46ed4056925072236994c31208f823f81 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_tensor([9993, 9993])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot surveys the bleak, sterile environment of the ward, bathed in the cool, flat light of overhead fluorescent lights. In the center of the frame, a pale-skinned individual with unkempt hair lies motionless on a metal-framed hospital bed. They are dressed in standard-issue light blue hospital gowns. With deliberate effort, the individual begins to slowly sit up; first their head lifts off the pillow, then their shoulders, using their elbows to elevate themselves. The crisp white cotton sheets, previously smooth and taut, now begin to gather and intricately wrinkle around the waist and legs, creating a complex textile topography. The movements are steady but tense, conveying a sense of physical frailty and quiet resolve. As they sit down, their backs slightly hunched, the camera captures their tired yet focused expression, capturing a profound moment of quiet in the otherwise hushed, clinical, and emotionally subdued atmosphere.", + "seed": 20994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20994_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13102273559197783, + "vlm_logprob": -0.13102273559197783, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.13102273559197783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5935847759246826 + ], + "opposite_logprob_tokens": [ + -0.8435847759246826 + ], + "logprob_reward": -0.5935847759246826, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5935847759246826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22494426369667053, + 0.04772523418068886, + 0.03568639978766441, + 0.08508306741714478, + 0.13096366822719574, + 0.12653644382953644, + 0.08686786890029907, + 0.0676322877407074, + 0.057315923273563385, + 0.13724488019943237 + ], + "credit_map": [ + 0.031660113483667374, + 0.027023974806070328, + 0.012747704982757568, + 0.024456309154629707, + 0.031122351065278053, + 0.02762848325073719, + 0.030938513576984406, + 0.039366818964481354, + 0.005637228488922119, + 0.0056749372743070126, + 0.0029471588786691427, + 0.005016105715185404, + 0.005249508190900087, + 0.005158049054443836, + 0.007386451121419668, + 0.010655795224010944, + 0.004581280052661896, + 0.004255523905158043, + 0.0024107845965772867, + 0.004503636155277491, + 0.005096036475151777, + 0.0044146208092570305, + 0.004465270321816206, + 0.005959250498563051, + 0.0099822748452425, + 0.014626868069171906, + 0.008547802455723286, + 0.008016378618776798, + 0.009768116287887096, + 0.01135034766048193, + 0.009514539502561092, + 0.013276739977300167, + 0.009512464515864849, + 0.016791125759482384, + 0.007892183028161526, + 0.014950600452721119, + 0.03318139165639877, + 0.02333845943212509, + 0.014630608260631561, + 0.010666826739907265, + 0.014637615531682968, + 0.024404199793934822, + 0.013055484741926193, + 0.013251936994493008, + 0.01731259562075138, + 0.017621038481593132, + 0.012274696491658688, + 0.013978865928947926, + 0.010192515328526497, + 0.011930015869438648, + 0.005296682473272085, + 0.014058561995625496, + 0.0130329430103302, + 0.011365964077413082, + 0.008703431114554405, + 0.01228775829076767, + 0.008302317932248116, + 0.011558512225747108, + 0.006774122826755047, + 0.009294586256146431, + 0.010074286721646786, + 0.009931406006217003, + 0.007382131181657314, + 0.004314921796321869, + 0.006803594995290041, + 0.007027489133179188, + 0.005425006151199341, + 0.013280240818858147, + 0.006139661651104689, + 0.008611713536083698, + 0.005780933424830437, + 0.004247286822646856, + 0.013990013860166073, + 0.016144638881087303, + 0.02550346963107586, + 0.012606480158865452, + 0.020528025925159454, + 0.019017497077584267, + 0.017837055027484894, + 0.011617699638009071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.5935847759246826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007661955431103706 + ], + "opposite_logprob_tokens": [ + -4.882661819458008 + ], + "logprob_reward": -0.007661955431103706, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007661955431103706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22541263699531555, + 0.08628684282302856, + 0.07516922801733017, + 0.08503914624452591, + 0.08682247251272202, + 0.06460355967283249, + 0.05717499181628227, + 0.11814575642347336, + 0.08876088261604309, + 0.11258444935083389 + ], + "credit_map": [ + 0.03251000493764877, + 0.024831825867295265, + 0.010973358526825905, + 0.022300321608781815, + 0.04064518213272095, + 0.03216595575213432, + 0.026061424985527992, + 0.035924553871154785, + 0.0075101712718605995, + 0.005278164055198431, + 0.004080073907971382, + 0.007617041934281588, + 0.007538881152868271, + 0.009457900188863277, + 0.01424736250191927, + 0.03055725246667862, + 0.008441286161541939, + 0.011669225059449673, + 0.0054290625266730785, + 0.00806793849915266, + 0.009920241311192513, + 0.009830786846578121, + 0.00953476969152689, + 0.01227591559290886, + 0.009183297865092754, + 0.013592069037258625, + 0.0063024647533893585, + 0.007391227409243584, + 0.010122815147042274, + 0.011464312672615051, + 0.010483356192708015, + 0.016499603167176247, + 0.0067351688630878925, + 0.013182345777750015, + 0.005544869229197502, + 0.009592604823410511, + 0.02125779539346695, + 0.015023741871118546, + 0.009733248502016068, + 0.005752692464739084, + 0.007150816265493631, + 0.011218409985303879, + 0.005867898464202881, + 0.008958744816482067, + 0.005252233240753412, + 0.00837685912847519, + 0.008002979680895805, + 0.009775620885193348, + 0.008526644669473171, + 0.010471025481820107, + 0.005035558715462685, + 0.009038728661835194, + 0.0064695593900978565, + 0.007933637127280235, + 0.004864812828600407, + 0.004835028667002916, + 0.014378014020621777, + 0.016804592683911324, + 0.01050720363855362, + 0.017058560624718666, + 0.01756388135254383, + 0.01898077130317688, + 0.014689384959638119, + 0.008163356222212315, + 0.010668283328413963, + 0.013999945484101772, + 0.006384831853210926, + 0.015216234140098095, + 0.014041420072317123, + 0.012549025937914848, + 0.007385337725281715, + 0.008515799418091774, + 0.012806057929992676, + 0.012212604284286499, + 0.027065334841609, + 0.01052535604685545, + 0.01274513453245163, + 0.014831854030489922, + 0.010989682748913765, + 0.011408424004912376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.007661955431103706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697523176670074 + ], + "opposite_logprob_tokens": [ + -3.4086976051330566 + ], + "logprob_reward": -0.033697523176670074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697523176670074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11879526078701019, + 0.047626569867134094, + 0.07508642226457596, + 0.0843164324760437, + 0.08104703575372696, + 0.10079912841320038, + 0.20604398846626282, + 0.05544668808579445, + 0.08699532598257065, + 0.1438431441783905 + ], + "credit_map": [ + 0.01409880816936493, + 0.013691660016775131, + 0.00697530061006546, + 0.021455243229866028, + 0.019097652286291122, + 0.015278877690434456, + 0.014704852364957333, + 0.013492868281900883, + 0.005163665860891342, + 0.00436093844473362, + 0.00353656685911119, + 0.007761246524751186, + 0.0037891583051532507, + 0.0048526497557759285, + 0.007603177335113287, + 0.01055916491895914, + 0.007969442754983902, + 0.010040467604994774, + 0.004577817860990763, + 0.007576202508062124, + 0.007609792053699493, + 0.008971061557531357, + 0.011837285943329334, + 0.016504349187016487, + 0.009321395307779312, + 0.007947868667542934, + 0.004845320712774992, + 0.008056385442614555, + 0.011678601615130901, + 0.010660405270755291, + 0.014677678234875202, + 0.017128780484199524, + 0.006819572299718857, + 0.012759453617036343, + 0.006941569037735462, + 0.013588812202215195, + 0.00756424106657505, + 0.0101211192086339, + 0.012597038410604, + 0.010655232705175877, + 0.013079802505671978, + 0.01662602834403515, + 0.007463887333869934, + 0.011372381821274757, + 0.011275752447545528, + 0.01577637530863285, + 0.01433507353067398, + 0.01086982898414135, + 0.029491335153579712, + 0.03176697716116905, + 0.01598636619746685, + 0.03514573350548744, + 0.024120057001709938, + 0.027949664741754532, + 0.01962621510028839, + 0.021957645192742348, + 0.006258592940866947, + 0.009090116247534752, + 0.005655727349221706, + 0.008585049770772457, + 0.008999352343380451, + 0.007656068541109562, + 0.00515186320990324, + 0.004049920942634344, + 0.011039897799491882, + 0.013747991994023323, + 0.007023216690868139, + 0.012119511142373085, + 0.014146325178444386, + 0.012639203108847141, + 0.008249737322330475, + 0.00802944041788578, + 0.01765196956694126, + 0.012283781543374062, + 0.031633779406547546, + 0.01322657335549593, + 0.021385785192251205, + 0.01876712031662464, + 0.017554238438606262, + 0.011339890770614147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.033697523176670074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098943900316954 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098943900316954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098943900316954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054138779640197754, + 0.039929069578647614, + 0.03695787489414215, + 0.04638165235519409, + 0.09226485341787338, + 0.030744269490242004, + 0.06880483776330948, + 0.16561752557754517, + 0.21281924843788147, + 0.2523418068885803 + ], + "credit_map": [ + 0.0063385204412043095, + 0.01060023345053196, + 0.003864868311211467, + 0.012144641019403934, + 0.004568018950521946, + 0.005601749289780855, + 0.00683597894385457, + 0.00418477039784193, + 0.00566776841878891, + 0.00613851984962821, + 0.0028303577564656734, + 0.007445828523486853, + 0.00416279211640358, + 0.004537141881883144, + 0.004658307880163193, + 0.004488356411457062, + 0.0031221574172377586, + 0.005790436640381813, + 0.0021921475417912006, + 0.0054492768831551075, + 0.004342260770499706, + 0.004603936802595854, + 0.006773997563868761, + 0.004683662671595812, + 0.006666133645921946, + 0.008763506077229977, + 0.003177434206008911, + 0.006477948743849993, + 0.003595472313463688, + 0.006037408020347357, + 0.007226080168038607, + 0.004437672905623913, + 0.00854423362761736, + 0.013480841182172298, + 0.005371255334466696, + 0.009307514876127243, + 0.023383069783449173, + 0.01711094193160534, + 0.008654410019516945, + 0.006412592250853777, + 0.0033924058079719543, + 0.006294170394539833, + 0.0024184021167457104, + 0.004123670980334282, + 0.003580070799216628, + 0.004331932868808508, + 0.003056934569031, + 0.003546682186424732, + 0.007580348756164312, + 0.013089357875287533, + 0.006241809576749802, + 0.012996428646147251, + 0.007220678497105837, + 0.010941664688289165, + 0.005450182594358921, + 0.005284375045448542, + 0.016724757850170135, + 0.01989107020199299, + 0.014715797267854214, + 0.03219129890203476, + 0.024052105844020844, + 0.022946037352085114, + 0.0226167943328619, + 0.01247966755181551, + 0.021752523258328438, + 0.03332706540822983, + 0.01501111313700676, + 0.028354983776807785, + 0.04976428300142288, + 0.03224722668528557, + 0.014879551716148853, + 0.017482509836554527, + 0.029150303453207016, + 0.030013982206583023, + 0.05314316973090172, + 0.01924549601972103, + 0.046093396842479706, + 0.03576085716485977, + 0.019057411700487137, + 0.019877204671502113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.004098943900316954 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070479527115822 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070479527115822, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070479527115822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056577980518341064, + 0.03367900848388672, + 0.02069266512989998, + 0.059756115078926086, + 0.08167004585266113, + 0.05289855971932411, + 0.05524664372205734, + 0.11546298116445541, + 0.1651432067155838, + 0.35887277126312256 + ], + "credit_map": [ + 0.005268971435725689, + 0.008605963550508022, + 0.003632857697084546, + 0.009023354388773441, + 0.008580437861382961, + 0.006478040013462305, + 0.007491580210626125, + 0.007496773265302181, + 0.0036960195284336805, + 0.004078234080225229, + 0.002267833100631833, + 0.004179715178906918, + 0.004239785950630903, + 0.0036664637736976147, + 0.004535782616585493, + 0.007015173323452473, + 0.002339510479941964, + 0.0029347194358706474, + 0.0014421268133446574, + 0.0028629912994802, + 0.003034032881259918, + 0.0027292734012007713, + 0.003225217806175351, + 0.002124794526025653, + 0.007506337948143482, + 0.010870049707591534, + 0.004094211850315332, + 0.008698885329067707, + 0.0042455680668354034, + 0.007945680059492588, + 0.009903358295559883, + 0.00649202661588788, + 0.006781314499676228, + 0.011467899195849895, + 0.004620356485247612, + 0.009263928048312664, + 0.02084208093583584, + 0.01514732837677002, + 0.008369498886168003, + 0.005177637562155724, + 0.0052588568069040775, + 0.009565337561070919, + 0.004762493073940277, + 0.00684355990961194, + 0.005287708714604378, + 0.006793868727982044, + 0.006297980900853872, + 0.008088752627372742, + 0.006939778570085764, + 0.008872512727975845, + 0.004819832276552916, + 0.00903753936290741, + 0.005332230124622583, + 0.006949308793991804, + 0.00658586947247386, + 0.006709570065140724, + 0.012757283635437489, + 0.01652780920267105, + 0.00966629944741726, + 0.021302498877048492, + 0.01184911746531725, + 0.016179967671632767, + 0.017253516241908073, + 0.00992649421095848, + 0.016696805134415627, + 0.028221214190125465, + 0.013262343592941761, + 0.02319587580859661, + 0.03413352742791176, + 0.022720206528902054, + 0.013344948180019855, + 0.013568301685154438, + 0.03595652058720589, + 0.03743855655193329, + 0.08656897395849228, + 0.027235310524702072, + 0.06997218728065491, + 0.050482116639614105, + 0.028615137562155724, + 0.022603990510106087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016070479527115822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot surveys the bleak, sterile environment of the ward, bathed in the cool, flat light of overhead fluorescent lights. In the center of the frame, a pale-skinned individual with unkempt hair lies motionless on a metal-framed hospital bed. They are dressed in standard-issue light blue hospital gowns. With deliberate effort, the individual begins to slowly sit up; first their head lifts off the pillow, then their shoulders, using their elbows to elevate themselves. The crisp white cotton sheets, previously smooth and taut, now begin to gather and intricately wrinkle around the waist and legs, creating a complex textile topography. The movements are steady but tense, conveying a sense of physical frailty and quiet resolve. As they sit down, their backs slightly hunched, the camera captures their tired yet focused expression, capturing a profound moment of quiet in the otherwise hushed, clinical, and emotionally subdued atmosphere.", + "seed": 20995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20995_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06559734996408224, + "vlm_logprob": -0.06559734996408224, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06559734996408224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21445100009441376 + ], + "opposite_logprob_tokens": [ + -1.8394509553909302 + ], + "logprob_reward": -0.21445100009441376, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21445100009441376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11802272498607635, + 0.13300056755542755, + 0.056293562054634094, + 0.1239888072013855, + 0.06181285157799721, + 0.13006125390529633, + 0.09869100898504257, + 0.16020168364048004, + 0.04717712104320526, + 0.07075038552284241 + ], + "credit_map": [ + 0.015664217993617058, + 0.019309913739562035, + 0.010758904740214348, + 0.01788552850484848, + 0.014213127084076405, + 0.013299659825861454, + 0.015122422017157078, + 0.01176895946264267, + 0.01423866767436266, + 0.01732364483177662, + 0.01061619445681572, + 0.019479481503367424, + 0.014099255204200745, + 0.016845272853970528, + 0.024668075144290924, + 0.015729984268546104, + 0.006941853556782007, + 0.008962051011621952, + 0.004383559804409742, + 0.007393485400825739, + 0.006275590043514967, + 0.008485925383865833, + 0.0074313427321612835, + 0.006419757381081581, + 0.01035863533616066, + 0.015002943575382233, + 0.013718907721340656, + 0.015812715515494347, + 0.023651795461773872, + 0.012369818985462189, + 0.021845946088433266, + 0.011228051036596298, + 0.005998796317726374, + 0.00788835622370243, + 0.005990167614072561, + 0.006645440123975277, + 0.01697522960603237, + 0.005347835831344128, + 0.007369925267994404, + 0.0055971043184399605, + 0.011131584644317627, + 0.010430676862597466, + 0.012069175019860268, + 0.027423998340964317, + 0.017265090718865395, + 0.01958860456943512, + 0.019429059699177742, + 0.012723075225949287, + 0.008189628832042217, + 0.008577744476497173, + 0.011318780481815338, + 0.017283489927649498, + 0.012175573967397213, + 0.00979333370923996, + 0.015654118731617928, + 0.015698343515396118, + 0.008787336759269238, + 0.014305539429187775, + 0.010036311112344265, + 0.03187243640422821, + 0.030037853866815567, + 0.01678096130490303, + 0.032358795404434204, + 0.0160224549472332, + 0.004926392808556557, + 0.006541248876601458, + 0.004660378210246563, + 0.006165639031678438, + 0.006400509271770716, + 0.005131176672875881, + 0.005333811976015568, + 0.008017966523766518, + 0.011818037368357182, + 0.011552629992365837, + 0.009919232688844204, + 0.01027145516127348, + 0.007372071035206318, + 0.00593110965564847, + 0.007062004413455725, + 0.006823853123933077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.21445100009441376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00597832165658474 + ], + "opposite_logprob_tokens": [ + -5.130978107452393 + ], + "logprob_reward": -0.00597832165658474, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00597832165658474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09826759248971939, + 0.06580359488725662, + 0.04431186988949776, + 0.09625992178916931, + 0.06759356707334518, + 0.07586301863193512, + 0.06634179502725601, + 0.16875551640987396, + 0.16639819741249084, + 0.1504049152135849 + ], + "credit_map": [ + 0.01000878307968378, + 0.014002485200762749, + 0.006081937346607447, + 0.015713438391685486, + 0.01422263216227293, + 0.013116014190018177, + 0.017280438914895058, + 0.007841868326067924, + 0.007522419560700655, + 0.00824388861656189, + 0.00629827007651329, + 0.008772415108978748, + 0.007897168397903442, + 0.008728801272809505, + 0.00975593738257885, + 0.008584695868194103, + 0.004430469591170549, + 0.008540001697838306, + 0.00359343527816236, + 0.006036903243511915, + 0.007113505620509386, + 0.004935871344059706, + 0.0059332046657800674, + 0.0037284826394170523, + 0.008863162249326706, + 0.010750972665846348, + 0.01168765127658844, + 0.012302658520638943, + 0.022630861029028893, + 0.00909661129117012, + 0.012711447663605213, + 0.0082165552303195, + 0.004571195226162672, + 0.007217723876237869, + 0.008130918256938457, + 0.009201697073876858, + 0.011419895105063915, + 0.005490089766681194, + 0.016166813671588898, + 0.00539523595944047, + 0.0056718760170042515, + 0.005495769903063774, + 0.007548038847744465, + 0.01555545162409544, + 0.008938989602029324, + 0.011436138302087784, + 0.013885092921555042, + 0.0073316665366292, + 0.005284902174025774, + 0.004604067653417587, + 0.0066253552213311195, + 0.011177060194313526, + 0.006706003565341234, + 0.005946509074419737, + 0.014175365678966045, + 0.01182253286242485, + 0.01189785823225975, + 0.013082911260426044, + 0.012131141498684883, + 0.03401505947113037, + 0.023476067930459976, + 0.016526993364095688, + 0.03002762235701084, + 0.02759786881506443, + 0.01971052587032318, + 0.022703876718878746, + 0.020933646708726883, + 0.02647240087389946, + 0.026715977117419243, + 0.01560983993113041, + 0.018337225541472435, + 0.01591470278799534, + 0.01177284400910139, + 0.021356159821152687, + 0.011338697746396065, + 0.03024369291961193, + 0.017532387748360634, + 0.007495037280023098, + 0.03675447404384613, + 0.013911630026996136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.00597832165658474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906334638595581 + ], + "opposite_logprob_tokens": [ + -2.4640634059906006 + ], + "logprob_reward": -0.08906334638595581, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906334638595581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08944573253393173, + 0.0469789020717144, + 0.07831363379955292, + 0.11403124034404755, + 0.10800410062074661, + 0.04366622120141983, + 0.05066082626581192, + 0.060775063931941986, + 0.10278639942407608, + 0.3053378760814667 + ], + "credit_map": [ + 0.008108295500278473, + 0.008784539997577667, + 0.009124292060732841, + 0.014509384520351887, + 0.010215417481958866, + 0.00895143672823906, + 0.022293705493211746, + 0.007458659820258617, + 0.005383372772485018, + 0.006133916322141886, + 0.004719370044767857, + 0.00843654852360487, + 0.0064238253980875015, + 0.005563510116189718, + 0.005588442552834749, + 0.004729917272925377, + 0.00667426735162735, + 0.009028775617480278, + 0.007832364179193974, + 0.014019057154655457, + 0.012194056063890457, + 0.010021119378507137, + 0.01117885671555996, + 0.007365129888057709, + 0.009907436557114124, + 0.012856281362473965, + 0.014168797992169857, + 0.016851909458637238, + 0.01395231019705534, + 0.014895395375788212, + 0.016809089109301567, + 0.01459001936018467, + 0.008088139817118645, + 0.01196710579097271, + 0.014032423496246338, + 0.012879983521997929, + 0.024741388857364655, + 0.009480491280555725, + 0.01438136026263237, + 0.012433216907083988, + 0.0045354170724749565, + 0.00570468045771122, + 0.00451108580455184, + 0.0077184224501252174, + 0.004905449692159891, + 0.006573513615876436, + 0.005092953331768513, + 0.004624694585800171, + 0.005897553637623787, + 0.005009440239518881, + 0.005917452275753021, + 0.007920707575976849, + 0.007048944476991892, + 0.006735502276569605, + 0.0075479778461158276, + 0.004583246074616909, + 0.0036037408281117678, + 0.007122690789401531, + 0.004149135667830706, + 0.009577763266861439, + 0.010087549686431885, + 0.005756953731179237, + 0.01181886252015829, + 0.008658363483846188, + 0.01382787898182869, + 0.014421944506466389, + 0.012977641075849533, + 0.015882933512330055, + 0.016635047271847725, + 0.009436975233256817, + 0.011814695782959461, + 0.007789277005940676, + 0.024939965456724167, + 0.031681645661592484, + 0.029788916930556297, + 0.06997976452112198, + 0.03302785009145737, + 0.019328448921442032, + 0.07270423322916031, + 0.023887058719992638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.08906334638595581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952606908977032 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952606908977032, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952606908977032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05636085197329521, + 0.022377746179699898, + 0.021236680448055267, + 0.039408858865499496, + 0.05726791173219681, + 0.06584436446428299, + 0.12702032923698425, + 0.1600760817527771, + 0.10032801330089569, + 0.3500792384147644 + ], + "credit_map": [ + 0.003750405739992857, + 0.007075851783156395, + 0.004945667460560799, + 0.009226791560649872, + 0.010337665677070618, + 0.006683078594505787, + 0.008322455920279026, + 0.006018929649144411, + 0.0020169394556432962, + 0.002584214089438319, + 0.0026666198391467333, + 0.004575975239276886, + 0.0023902510292828083, + 0.0022785086184740067, + 0.0033049951307475567, + 0.0025602413807064295, + 0.0015546869253739715, + 0.003529188921675086, + 0.0014434918994084, + 0.0034338384866714478, + 0.0033537731505930424, + 0.002813800470903516, + 0.0029332288540899754, + 0.0021746691782027483, + 0.0025893517304211855, + 0.003677159082144499, + 0.0037840791046619415, + 0.008330663666129112, + 0.004128587432205677, + 0.003842428093776107, + 0.00986261386424303, + 0.0031939707696437836, + 0.005970625672489405, + 0.006014613900333643, + 0.0073068346828222275, + 0.0077946921810507774, + 0.009969829581677914, + 0.005689895711839199, + 0.008495100773870945, + 0.006026314105838537, + 0.005502248648554087, + 0.004952579736709595, + 0.007424923591315746, + 0.010050847195088863, + 0.008103187195956707, + 0.005521021783351898, + 0.016957681626081467, + 0.007331869099289179, + 0.00952185969799757, + 0.01122873555868864, + 0.010298245586454868, + 0.0184787530452013, + 0.021989667788147926, + 0.021608417853713036, + 0.020019788295030594, + 0.013874854892492294, + 0.008265086449682713, + 0.01418247353285551, + 0.010289612226188183, + 0.03033415786921978, + 0.02384003810584545, + 0.020086543634533882, + 0.025757644325494766, + 0.0273205004632473, + 0.012037016451358795, + 0.012998631224036217, + 0.01223880983889103, + 0.011515103280544281, + 0.013553289696574211, + 0.010730003006756306, + 0.015600588172674179, + 0.011654563248157501, + 0.023269904777407646, + 0.05567757785320282, + 0.027031486853957176, + 0.06229071691632271, + 0.04497905075550079, + 0.025294311344623566, + 0.06848622858524323, + 0.04304992035031319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.005952606908977032 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541474774479866 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541474774479866, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541474774479866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06842994689941406, + 0.06617999076843262, + 0.02959730476140976, + 0.11606110632419586, + 0.06316597014665604, + 0.12185873091220856, + 0.13944599032402039, + 0.08244884014129639, + 0.17453141510486603, + 0.1382807195186615 + ], + "credit_map": [ + 0.008056185208261013, + 0.010090663097798824, + 0.009420789778232574, + 0.011697161011397839, + 0.006687825080007315, + 0.00824099499732256, + 0.008538697846233845, + 0.005697628948837519, + 0.006987494416534901, + 0.008393540047109127, + 0.004217337816953659, + 0.009727501310408115, + 0.006883406080305576, + 0.008522159419953823, + 0.013572420924901962, + 0.007876127026975155, + 0.0025044800713658333, + 0.003787691704928875, + 0.0023548807948827744, + 0.0041388049721717834, + 0.004763690754771233, + 0.0033064705785363913, + 0.004944771062582731, + 0.0037965159863233566, + 0.010643615387380123, + 0.013641762547194958, + 0.01321711577475071, + 0.013352148234844208, + 0.02441725879907608, + 0.012925663962960243, + 0.016438856720924377, + 0.011424684897065163, + 0.005467851180583239, + 0.007299432530999184, + 0.007279304321855307, + 0.010150842368602753, + 0.008133777417242527, + 0.005914920940995216, + 0.009642304852604866, + 0.009277540259063244, + 0.009843278676271439, + 0.009184499271214008, + 0.01169538963586092, + 0.01848769374191761, + 0.01749551296234131, + 0.02149093896150589, + 0.019492141902446747, + 0.01416927669197321, + 0.018414195626974106, + 0.01045086421072483, + 0.018962308764457703, + 0.018356459215283394, + 0.019974926486611366, + 0.016536695882678032, + 0.01931121200323105, + 0.017439335584640503, + 0.007954627275466919, + 0.008841916918754578, + 0.008653988130390644, + 0.013291280716657639, + 0.011068766005337238, + 0.010541962459683418, + 0.011368920095264912, + 0.010727380402386189, + 0.023956002667546272, + 0.020977791398763657, + 0.02933536097407341, + 0.026069119572639465, + 0.02528269961476326, + 0.012870767153799534, + 0.02143982984125614, + 0.014599842019379139, + 0.020705392584204674, + 0.021722562611103058, + 0.016106242313981056, + 0.019758325070142746, + 0.015832582488656044, + 0.01618105359375477, + 0.014587818644940853, + 0.01338675431907177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.012541474774479866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot surveys the bleak, sterile environment of the ward, bathed in the cool, flat light of overhead fluorescent lights. In the center of the frame, a pale-skinned individual with unkempt hair lies motionless on a metal-framed hospital bed. They are dressed in standard-issue light blue hospital gowns. With deliberate effort, the individual begins to slowly sit up; first their head lifts off the pillow, then their shoulders, using their elbows to elevate themselves. The crisp white cotton sheets, previously smooth and taut, now begin to gather and intricately wrinkle around the waist and legs, creating a complex textile topography. The movements are steady but tense, conveying a sense of physical frailty and quiet resolve. As they sit down, their backs slightly hunched, the camera captures their tired yet focused expression, capturing a profound moment of quiet in the otherwise hushed, clinical, and emotionally subdued atmosphere.", + "seed": 20996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20996_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08222852339968086, + "vlm_logprob": -0.08222852339968086, + "vlm_yes_no_margin": 4.025000011920929, + "vlm_token_credit": -0.08222852339968086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3679894804954529 + ], + "opposite_logprob_tokens": [ + -1.2429895401000977 + ], + "logprob_reward": -0.3679894804954529, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3679894804954529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1385279893875122, + 0.05398533493280411, + 0.038711413741111755, + 0.033952292054891586, + 0.04889889806509018, + 0.05231422185897827, + 0.047530654817819595, + 0.08020665496587753, + 0.2602669894695282, + 0.24560551345348358 + ], + "credit_map": [ + 0.01049021165817976, + 0.014782154932618141, + 0.011442878283560276, + 0.024462303146719933, + 0.018938686698675156, + 0.010542439296841621, + 0.026883067563176155, + 0.020986245945096016, + 0.0045185936614871025, + 0.005063179414719343, + 0.004049560986459255, + 0.009421507827937603, + 0.005847457330673933, + 0.004622000735253096, + 0.011671659536659718, + 0.00879137497395277, + 0.005236828234046698, + 0.0031763033475726843, + 0.0037319615948945284, + 0.00564020499587059, + 0.00487635750323534, + 0.0029680158477276564, + 0.0072390311397612095, + 0.005842713173478842, + 0.0030392110347747803, + 0.0031753613147884607, + 0.0030854246579110622, + 0.004311056807637215, + 0.0056425221264362335, + 0.004537872038781643, + 0.005353836342692375, + 0.004807009361684322, + 0.008005483075976372, + 0.00395288085564971, + 0.004696432966738939, + 0.006744868587702513, + 0.006540987640619278, + 0.005514818709343672, + 0.006503553129732609, + 0.006939878687262535, + 0.005512057337909937, + 0.007500207517296076, + 0.00557874096557498, + 0.007317046169191599, + 0.0060045961290597916, + 0.007180478889495134, + 0.007435400504618883, + 0.005785695277154446, + 0.007347981445491314, + 0.005195931531488895, + 0.003287020605057478, + 0.00548891955986619, + 0.005370922852307558, + 0.004389481619000435, + 0.0103314109146595, + 0.006118991412222385, + 0.007759974338114262, + 0.008809378370642662, + 0.00750560499727726, + 0.013008525595068932, + 0.010099981911480427, + 0.014672539196908474, + 0.007545288186520338, + 0.01080536562949419, + 0.0539446622133255, + 0.020900750532746315, + 0.020895875990390778, + 0.015881896018981934, + 0.019199255853891373, + 0.04573388770222664, + 0.05412543565034866, + 0.029585259035229683, + 0.030791401863098145, + 0.023104170337319374, + 0.03057452104985714, + 0.026431208476424217, + 0.01764676161110401, + 0.03910720720887184, + 0.04954328015446663, + 0.028406957164406776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.3679894804954529 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004118651617318392 + ], + "opposite_logprob_tokens": [ + -5.5041184425354 + ], + "logprob_reward": -0.004118651617318392, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004118651617318392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20751851797103882, + 0.06531717628240585, + 0.05350780114531517, + 0.03415469452738762, + 0.10311073809862137, + 0.047684092074632645, + 0.029017627239227295, + 0.07424900680780411, + 0.17692355811595917, + 0.20851679146289825 + ], + "credit_map": [ + 0.029716093093156815, + 0.021023930981755257, + 0.016247427091002464, + 0.03458278253674507, + 0.028872888535261154, + 0.014806103892624378, + 0.03684358671307564, + 0.02542569674551487, + 0.00732199614867568, + 0.006356007419526577, + 0.0049942792393267155, + 0.013085944578051567, + 0.007385837845504284, + 0.004613544326275587, + 0.010780422016978264, + 0.010779145173728466, + 0.0048728203400969505, + 0.005783120635896921, + 0.004335668403655291, + 0.008792162872850895, + 0.007838222198188305, + 0.004368020687252283, + 0.009469146840274334, + 0.00804863777011633, + 0.0035557171795517206, + 0.0021578555461019278, + 0.0027125130873173475, + 0.0036556886043399572, + 0.007094650994986296, + 0.004239395726472139, + 0.005313225090503693, + 0.005425648298114538, + 0.013119435869157314, + 0.00772223761305213, + 0.008884906768798828, + 0.016672426834702492, + 0.013783660717308521, + 0.01139208022505045, + 0.016247043386101723, + 0.015288950875401497, + 0.005119437351822853, + 0.007351957727223635, + 0.004572920501232147, + 0.007148699834942818, + 0.007448645308613777, + 0.004611797630786896, + 0.006064873654395342, + 0.005365758202970028, + 0.004122445825487375, + 0.004609591327607632, + 0.0018250419525429606, + 0.00268327328376472, + 0.003245576983317733, + 0.0029137455858290195, + 0.005846983753144741, + 0.003770969808101654, + 0.007575471419841051, + 0.009934548288583755, + 0.007851962931454182, + 0.011930020526051521, + 0.011385684832930565, + 0.009583687409758568, + 0.007635178975760937, + 0.008352456614375114, + 0.037173639982938766, + 0.013175678439438343, + 0.01427585631608963, + 0.010747665539383888, + 0.013200686313211918, + 0.02989577129483223, + 0.037865713238716125, + 0.020588546991348267, + 0.028299780562520027, + 0.020587293431162834, + 0.027632083743810654, + 0.025922518223524094, + 0.014958670362830162, + 0.02671273984014988, + 0.04222304746508598, + 0.022180672734975815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004118651617318392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741754245012999 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741754245012999, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741754245012999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07669274508953094, + 0.040136225521564484, + 0.055954474955797195, + 0.042623370885849, + 0.033551376312971115, + 0.04889342188835144, + 0.05049045383930206, + 0.1674259752035141, + 0.14795321226119995, + 0.3362787663936615 + ], + "credit_map": [ + 0.007879146374762058, + 0.0074362987652421, + 0.005117751657962799, + 0.013900207355618477, + 0.012608152814209461, + 0.005302734207361937, + 0.014900698326528072, + 0.009547747671604156, + 0.003936913330107927, + 0.003469966584816575, + 0.002812211401760578, + 0.0067512765526771545, + 0.004296541213989258, + 0.002654632553458214, + 0.009587588720023632, + 0.0066270893439650536, + 0.011110694147646427, + 0.0029790475964546204, + 0.004917835351079702, + 0.008499660529196262, + 0.00519412849098444, + 0.005287598352879286, + 0.00994334276765585, + 0.008022163063287735, + 0.006225044373422861, + 0.003392291022464633, + 0.004333260003477335, + 0.006041446700692177, + 0.006060010753571987, + 0.004926328081637621, + 0.0061998735181987286, + 0.00544511154294014, + 0.00433883722871542, + 0.003331247717142105, + 0.0027165364008396864, + 0.005278578028082848, + 0.005132108461111784, + 0.0038523762486875057, + 0.004200515802949667, + 0.004701175726950169, + 0.004465932492166758, + 0.008043845184147358, + 0.00322410324588418, + 0.007328530307859182, + 0.006375242490321398, + 0.003807758679613471, + 0.009786766953766346, + 0.005861238110810518, + 0.008037025108933449, + 0.0054281000047922134, + 0.0036228199023753405, + 0.0066985152661800385, + 0.005384562537074089, + 0.006532381754368544, + 0.008654179982841015, + 0.006132862064987421, + 0.015509005635976791, + 0.02133067511022091, + 0.015899397432804108, + 0.023018086329102516, + 0.026906706392765045, + 0.026612967252731323, + 0.01718413457274437, + 0.0209649745374918, + 0.035589221864938736, + 0.011689593084156513, + 0.01173039898276329, + 0.00729329651221633, + 0.014828665181994438, + 0.019231846556067467, + 0.03751072287559509, + 0.01007943507283926, + 0.048162199556827545, + 0.045813679695129395, + 0.048331499099731445, + 0.04587460309267044, + 0.029806863516569138, + 0.03367273882031441, + 0.056517936289310455, + 0.02809920720756054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.006741754245012999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950947757810354 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950947757810354, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950947757810354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14707674086093903, + 0.06888341903686523, + 0.07849690318107605, + 0.08424898982048035, + 0.09808333963155746, + 0.07802087813615799, + 0.08632354438304901, + 0.08345407247543335, + 0.20446406304836273, + 0.07094797492027283 + ], + "credit_map": [ + 0.014998095110058784, + 0.014107413589954376, + 0.014477158896625042, + 0.024824731051921844, + 0.014427945017814636, + 0.013602920807898045, + 0.026570184156298637, + 0.024068299680948257, + 0.007152629550546408, + 0.007531354203820229, + 0.005510833580046892, + 0.01017374824732542, + 0.007430346682667732, + 0.006396028678864241, + 0.01429151464253664, + 0.010396965779364109, + 0.007804912514984608, + 0.006252909079194069, + 0.006235884502530098, + 0.014468635432422161, + 0.0065274424850940704, + 0.007957210764288902, + 0.01632797159254551, + 0.012921929359436035, + 0.007758787367492914, + 0.007853895425796509, + 0.006734957918524742, + 0.01255564484745264, + 0.01768721267580986, + 0.009984450414776802, + 0.01060875691473484, + 0.01106528751552105, + 0.013204113580286503, + 0.007896743714809418, + 0.010751239955425262, + 0.012051683850586414, + 0.015044736675918102, + 0.008628468960523605, + 0.01588205061852932, + 0.014624310657382011, + 0.009223029948771, + 0.009282675571739674, + 0.006305666174739599, + 0.010939578525722027, + 0.009460676461458206, + 0.011977793648838997, + 0.012139174155890942, + 0.008692283183336258, + 0.014135710895061493, + 0.009174133650958538, + 0.007706924807280302, + 0.011068366467952728, + 0.00984685868024826, + 0.008747072890400887, + 0.01563786342740059, + 0.010006613098084927, + 0.007589309010654688, + 0.013491569086909294, + 0.008496960625052452, + 0.009579453617334366, + 0.013841625303030014, + 0.009487021714448929, + 0.009182726964354515, + 0.01178540289402008, + 0.04767554998397827, + 0.01932789757847786, + 0.021587297320365906, + 0.013353132642805576, + 0.01950434409081936, + 0.014680149964988232, + 0.05269216001033783, + 0.015643537044525146, + 0.00862968061119318, + 0.007827876135706902, + 0.0058994293212890625, + 0.008563141338527203, + 0.006478444207459688, + 0.01100208517163992, + 0.015283613465726376, + 0.007263696752488613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.005950947757810354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634178288280964 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.02634178288280964, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634178288280964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1801794022321701, + 0.054445985704660416, + 0.04241923615336418, + 0.05944298580288887, + 0.08370250463485718, + 0.12286174297332764, + 0.05280312895774841, + 0.15794572234153748, + 0.1820482313632965, + 0.06415096670389175 + ], + "credit_map": [ + 0.03835805132985115, + 0.013891697861254215, + 0.019892113283276558, + 0.025059819221496582, + 0.0200706385076046, + 0.016772985458374023, + 0.025044510141015053, + 0.021089604124426842, + 0.010520879179239273, + 0.004091762471944094, + 0.005650566425174475, + 0.007948973216116428, + 0.005524190608412027, + 0.0047186510637402534, + 0.008708225563168526, + 0.007282748352736235, + 0.0035542771220207214, + 0.003481591120362282, + 0.003943002782762051, + 0.006064021959900856, + 0.004574708640575409, + 0.004852237645536661, + 0.008508170023560524, + 0.007441231515258551, + 0.010122734121978283, + 0.004641150590032339, + 0.007215189281851053, + 0.00904928334057331, + 0.007012973073869944, + 0.00555985514074564, + 0.007357683498412371, + 0.008484123274683952, + 0.0093077402561903, + 0.007283552549779415, + 0.007728587836027145, + 0.016719400882720947, + 0.009319497272372246, + 0.009252113290131092, + 0.011572222225368023, + 0.012519403360784054, + 0.016123734414577484, + 0.02805485762655735, + 0.009319552220404148, + 0.01745220273733139, + 0.014235645532608032, + 0.010465609841048717, + 0.01217935886234045, + 0.01503080502152443, + 0.00917886570096016, + 0.006542070768773556, + 0.003925163298845291, + 0.006530978251248598, + 0.005057030823081732, + 0.003833804512396455, + 0.0110967131331563, + 0.006638512015342712, + 0.014377456158399582, + 0.023078924044966698, + 0.014364470727741718, + 0.024302076548337936, + 0.023823663592338562, + 0.018412910401821136, + 0.01689504273235798, + 0.02269120328128338, + 0.04531130939722061, + 0.013288978487253189, + 0.013367338106036186, + 0.008654394187033176, + 0.012365666218101978, + 0.02287546917796135, + 0.04870135337114334, + 0.017483746632933617, + 0.008910875767469406, + 0.006601367611438036, + 0.007966962642967701, + 0.007584940642118454, + 0.005559801589697599, + 0.008152386173605919, + 0.012752917595207691, + 0.006621724925935268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02634178288280964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot surveys the bleak, sterile environment of the ward, bathed in the cool, flat light of overhead fluorescent lights. In the center of the frame, a pale-skinned individual with unkempt hair lies motionless on a metal-framed hospital bed. They are dressed in standard-issue light blue hospital gowns. With deliberate effort, the individual begins to slowly sit up; first their head lifts off the pillow, then their shoulders, using their elbows to elevate themselves. The crisp white cotton sheets, previously smooth and taut, now begin to gather and intricately wrinkle around the waist and legs, creating a complex textile topography. The movements are steady but tense, conveying a sense of physical frailty and quiet resolve. As they sit down, their backs slightly hunched, the camera captures their tired yet focused expression, capturing a profound moment of quiet in the otherwise hushed, clinical, and emotionally subdued atmosphere.", + "seed": 20997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_7/index_9993_seed_20997_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018581591080874205, + "vlm_logprob": -0.018581591080874205, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.018581591080874205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04437577724456787 + ], + "opposite_logprob_tokens": [ + -3.6693758964538574 + ], + "logprob_reward": -0.04437577724456787, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04437577724456787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21654008328914642, + 0.017982127144932747, + 0.01865210197865963, + 0.30816975235939026, + 0.04457563906908035, + 0.0678681954741478, + 0.19229070842266083, + 0.035607945173978806, + 0.039666108787059784, + 0.05864734202623367 + ], + "credit_map": [ + 0.024149153381586075, + 0.021657759323716164, + 0.03455393761396408, + 0.025428924709558487, + 0.024500489234924316, + 0.02665764093399048, + 0.0319158099591732, + 0.027676355093717575, + 0.0022162687964737415, + 0.002445284044370055, + 0.0012852143263444304, + 0.0022846029605716467, + 0.00386248342692852, + 0.0019448179518803954, + 0.001989287557080388, + 0.001954165520146489, + 0.0024948616046458483, + 0.002324968110769987, + 0.0013070538407191634, + 0.00243168231099844, + 0.0027685079257935286, + 0.00268100225366652, + 0.0024950355291366577, + 0.002148988423869014, + 0.008493350818753242, + 0.011120284907519817, + 0.007264349609613419, + 0.13495314121246338, + 0.008845595642924309, + 0.014861254021525383, + 0.1084270179271698, + 0.014204747043550014, + 0.0043357498943805695, + 0.004250984638929367, + 0.005965667776763439, + 0.007621485739946365, + 0.005632216110825539, + 0.0037300013937056065, + 0.008778606541454792, + 0.004260924179106951, + 0.005772363394498825, + 0.008086031302809715, + 0.002503695199266076, + 0.013827132061123848, + 0.009463664144277573, + 0.006456949282437563, + 0.013795028440654278, + 0.007963327690958977, + 0.028439141809940338, + 0.029392754659056664, + 0.016776911914348602, + 0.0202187392860651, + 0.038985177874565125, + 0.0231925081461668, + 0.01928061619400978, + 0.01600484549999237, + 0.004264847841113806, + 0.007225012872368097, + 0.005073847249150276, + 0.0038399070035666227, + 0.00411496264860034, + 0.004708317574113607, + 0.003162417793646455, + 0.0032186307944357395, + 0.0034056315198540688, + 0.005365968216210604, + 0.0027465191669762135, + 0.00775194214656949, + 0.004259718116372824, + 0.004082880448549986, + 0.007492018397897482, + 0.0045614284463226795, + 0.004550632555037737, + 0.00441618962213397, + 0.003476229961961508, + 0.00774648180231452, + 0.004725179634988308, + 0.014227679930627346, + 0.012847332283854485, + 0.006657615303993225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.04437577724456787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001962760929018259 + ], + "opposite_logprob_tokens": [ + -6.251962661743164 + ], + "logprob_reward": -0.001962760929018259, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001962760929018259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14946609735488892, + 0.1757519692182541, + 0.018695946782827377, + 0.0773572027683258, + 0.07150880247354507, + 0.13650108873844147, + 0.194912388920784, + 0.041212230920791626, + 0.05122637003660202, + 0.08336786180734634 + ], + "credit_map": [ + 0.017779862508177757, + 0.01578044705092907, + 0.03196040168404579, + 0.01684633642435074, + 0.013212415389716625, + 0.018267320469021797, + 0.017934901639819145, + 0.017684420570731163, + 0.0224513690918684, + 0.01880953460931778, + 0.026979470625519753, + 0.02422984503209591, + 0.02043423429131508, + 0.018609657883644104, + 0.02160348929464817, + 0.02263438142836094, + 0.0025638537481427193, + 0.0023979905527085066, + 0.0009325612336397171, + 0.002706528641283512, + 0.003471147734671831, + 0.0023979938123375177, + 0.0021314993500709534, + 0.0020943726412951946, + 0.005029615946114063, + 0.005623775068670511, + 0.010351106524467468, + 0.006171712186187506, + 0.006365418899804354, + 0.005618147552013397, + 0.01590788923203945, + 0.022289549931883812, + 0.0060691810213029385, + 0.008124108426272869, + 0.004921788349747658, + 0.013624339364469051, + 0.008631750009953976, + 0.006137270946055651, + 0.016336552798748016, + 0.007663820870220661, + 0.009866288863122463, + 0.009217169135808945, + 0.0046304818242788315, + 0.039685823023319244, + 0.013355673290789127, + 0.009417060762643814, + 0.041587285697460175, + 0.008741309866309166, + 0.02451130375266075, + 0.027360672131180763, + 0.017762228846549988, + 0.03047369047999382, + 0.032756175845861435, + 0.020525678992271423, + 0.022728828713297844, + 0.01879383623600006, + 0.006726272404193878, + 0.007857373915612698, + 0.004430747590959072, + 0.004510022699832916, + 0.004370608367025852, + 0.00561740156263113, + 0.004475134890526533, + 0.0032246753107756376, + 0.005126440431922674, + 0.007694692350924015, + 0.004205906298011541, + 0.00762690557166934, + 0.006832856684923172, + 0.006086201872676611, + 0.007871448993682861, + 0.00578191876411438, + 0.00616456801071763, + 0.0071283006109297276, + 0.005482897628098726, + 0.012512879446148872, + 0.005914653651416302, + 0.014966774731874466, + 0.02130417339503765, + 0.009893619455397129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.001962760929018259 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763084227219224 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.00763084227219224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763084227219224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04397640749812126, + 0.07258680462837219, + 0.018390394747257233, + 0.3557659983634949, + 0.11207418888807297, + 0.09953340142965317, + 0.07896661013364792, + 0.07683906704187393, + 0.055451277643442154, + 0.08641581982374191 + ], + "credit_map": [ + 0.004819225054234266, + 0.0050066085532307625, + 0.0043548038229346275, + 0.005699200555682182, + 0.007068423554301262, + 0.005252690054476261, + 0.004710361827164888, + 0.007065092213451862, + 0.009870660491287708, + 0.007541227154433727, + 0.011761887930333614, + 0.008924776688218117, + 0.008791507221758366, + 0.007551664020866156, + 0.009240546263754368, + 0.008904539979994297, + 0.0020178817212581635, + 0.002225474687293172, + 0.0020055309869349003, + 0.002809843746945262, + 0.0023906847927719355, + 0.002233442384749651, + 0.0021349340677261353, + 0.002572603989392519, + 0.00942399725317955, + 0.012790976092219353, + 0.009627924300730228, + 0.1548926830291748, + 0.0110316202044487, + 0.015388106927275658, + 0.1222057119011879, + 0.02040497586131096, + 0.006176464259624481, + 0.009282422251999378, + 0.0035714718978852034, + 0.03238627314567566, + 0.01062578521668911, + 0.007449118420481682, + 0.03268225118517876, + 0.009900400415062904, + 0.010448645800352097, + 0.01213906891644001, + 0.0059684826992452145, + 0.01866157539188862, + 0.011908738873898983, + 0.010166547261178493, + 0.0176906306296587, + 0.012549711391329765, + 0.010562422685325146, + 0.01063152402639389, + 0.005136987194418907, + 0.010581331327557564, + 0.014465944841504097, + 0.010725879110395908, + 0.010381991975009441, + 0.0064805252477526665, + 0.01094578392803669, + 0.013543244451284409, + 0.007914133369922638, + 0.008613137528300285, + 0.011830831877887249, + 0.009521151892840862, + 0.007402187678962946, + 0.007068594451993704, + 0.004811285063624382, + 0.007636223919689655, + 0.003624655306339264, + 0.006070415023714304, + 0.007724621798843145, + 0.006328905466943979, + 0.0072382413782179356, + 0.01201692782342434, + 0.006348246708512306, + 0.007471400313079357, + 0.006433654110878706, + 0.013500860892236233, + 0.005728425458073616, + 0.013742439448833466, + 0.023507436737418175, + 0.009683351963758469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.00763084227219224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251422990113497 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251422990113497, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251422990113497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039343494921922684, + 0.016339581459760666, + 0.01348021812736988, + 0.44330334663391113, + 0.07010935246944427, + 0.13976292312145233, + 0.18086662888526917, + 0.0191651601344347, + 0.029917795211076736, + 0.04771146923303604 + ], + "credit_map": [ + 0.004469532053917646, + 0.004483397118747234, + 0.0059760091826319695, + 0.004761949647217989, + 0.005487944930791855, + 0.0045861005783081055, + 0.004034815356135368, + 0.005543744657188654, + 0.0023165603633970022, + 0.0022657576482743025, + 0.002019529463723302, + 0.0017805593088269234, + 0.0028098891489207745, + 0.0016342677408829331, + 0.0016592945903539658, + 0.0018537224968895316, + 0.0013552968157455325, + 0.001249635242857039, + 0.0016072560101747513, + 0.0018498161807656288, + 0.0016237142262980342, + 0.0019540819339454174, + 0.001754173543304205, + 0.0020862435922026634, + 0.011786898598074913, + 0.015537993051111698, + 0.009935406036674976, + 0.19468094408512115, + 0.013497771695256233, + 0.019244279712438583, + 0.1510794311761856, + 0.027540626004338264, + 0.003625282319262624, + 0.005313403904438019, + 0.0031096008606255054, + 0.020985238254070282, + 0.007140554953366518, + 0.003972588572651148, + 0.020447317510843277, + 0.005515367724001408, + 0.007992701604962349, + 0.01004035770893097, + 0.004911254625767469, + 0.044610291719436646, + 0.009484349749982357, + 0.007649586535990238, + 0.047003235667943954, + 0.008071144111454487, + 0.024687618017196655, + 0.02803184650838375, + 0.013921036384999752, + 0.020549744367599487, + 0.03823193907737732, + 0.02175099588930607, + 0.019407643005251884, + 0.014285811223089695, + 0.002709317021071911, + 0.0019983150996267796, + 0.0033154450356960297, + 0.002480096882209182, + 0.0016225004801526666, + 0.0021610278636217117, + 0.0026259932201355696, + 0.002252466045320034, + 0.0032230359502136707, + 0.004502235911786556, + 0.0021263703238219023, + 0.003532120492309332, + 0.00397596275433898, + 0.004552710801362991, + 0.003860956523567438, + 0.004144403152167797, + 0.003280433127656579, + 0.0038857152685523033, + 0.004222879186272621, + 0.007690805941820145, + 0.0027111261151731014, + 0.0058461143635213375, + 0.014282725751399994, + 0.005791671574115753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.005251422990113497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile hospital ward with cool, flat fluorescent lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale-skinned individual with unkempt hair lying on a metal-framed hospital bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the individual wearing a standard-issue light blue hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the individual slowly sit up using their elbows, causing the white cotton sheets to wrinkle around their waist and legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687151968479156 + ], + "opposite_logprob_tokens": [ + -3.408687114715576 + ], + "logprob_reward": -0.033687151968479156, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687151968479156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023859268054366112, + 0.01570410467684269, + 0.01214244868606329, + 0.6812041401863098, + 0.035774245858192444, + 0.07123931497335434, + 0.05688144639134407, + 0.018238220363855362, + 0.030077187344431877, + 0.054879575967788696 + ], + "credit_map": [ + 0.002799462992697954, + 0.0033572090324014425, + 0.0015010653296485543, + 0.003236922202631831, + 0.004002206958830357, + 0.0032276869751513004, + 0.0025128633715212345, + 0.0032218548003584146, + 0.0023319392930716276, + 0.0019314539385959506, + 0.0012526096543297172, + 0.0021172715350985527, + 0.0024647535756230354, + 0.001753646065481007, + 0.0020390902645885944, + 0.0018133417470380664, + 0.0016075451858341694, + 0.0016167891444638371, + 0.0015025417087599635, + 0.0013175054918974638, + 0.0016858285525813699, + 0.0017528270836919546, + 0.0013628380838781595, + 0.0012965751811861992, + 0.01804969273507595, + 0.023807331919670105, + 0.015616956166923046, + 0.2993101477622986, + 0.020043712109327316, + 0.030151300132274628, + 0.23322629928588867, + 0.04099881649017334, + 0.0031839655712246895, + 0.0030118627473711967, + 0.0027532761450856924, + 0.007514354307204485, + 0.00402647303417325, + 0.0031016278080642223, + 0.008171367458999157, + 0.004011321812868118, + 0.0037778010591864586, + 0.004356182646006346, + 0.0017632823437452316, + 0.024951735511422157, + 0.0033643487840890884, + 0.004118555691093206, + 0.025745008140802383, + 0.0031624087132513523, + 0.009165903553366661, + 0.008210846222937107, + 0.0055784485302865505, + 0.005100777838379145, + 0.0111979516223073, + 0.007003394421190023, + 0.005310985259711742, + 0.005313141271471977, + 0.0024823632556945086, + 0.002594962250441313, + 0.0030818390659987926, + 0.0019125822000205517, + 0.0022898968309164047, + 0.0021838194224983454, + 0.0017519189277663827, + 0.001940839341841638, + 0.0028107997495681047, + 0.004390880931168795, + 0.0031473515555262566, + 0.003837455762550235, + 0.0028524629306048155, + 0.004518796224147081, + 0.005407201591879129, + 0.0031122418586164713, + 0.005859408061951399, + 0.004881291184574366, + 0.003893494140356779, + 0.007839460857212543, + 0.005326039157807827, + 0.012894853949546814, + 0.00873487163335085, + 0.005450164433568716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.033687151968479156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_tensor([9229, 9229])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_tensor([9229, 9229])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..91d7285198d127af7372064b702b70b7d1351323 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_tensor([9229, 9229])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up shot framed in a vertical format, the camera tilts steadily downward from a slight downward angle, following a brass pocket watch, intricately engraved with vintage patterns, as it slowly slides down the surface of a tilted, deep burgundy silk handkerchief. The aged brass case, softly lit from the side, gleams with a warm, deep metallic luster. The interplay of light and shadow lends the surface a vivid three-dimensional quality, while the slight wear on the edges betrays the wear of time. The delicate texture of the silk handkerchief is clearly visible, its satiny sheen shifting and changing with the movement of the watch. The watch's descent is not sudden, but rather steady and deliberate, its weight indenting the soft silk with a shallow, shifting indentation. The camera precisely captures every subtle lag between metal and fabric, almost perceiving the silent resistance. The entire picture is shrouded in a tranquil and nostalgic atmosphere. The high level of detail captured creates a sharp contrast between the cold hardness of brass and the soft warmth of silk, creating an elegant meditation on the passage of time and the beauty of material things.", + "seed": 2478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_9229_seed_2478_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19877261598594487, + "vlm_logprob": -0.19877261598594487, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.19877261598594487, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07554557174444199 + ], + "opposite_logprob_tokens": [ + -3.3255455493927 + ], + "logprob_reward": -0.07554557174444199, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.07554557174444199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10267752408981323, + 0.04172055423259735, + 0.07559295743703842, + 0.034750234335660934, + 0.663614809513092, + 0.0213541891425848, + 0.013232104480266571, + 0.015660064294934273, + 0.018421167507767677, + 0.012976438738405704 + ], + "credit_map": [ + 0.011275634169578552, + 0.013640317134559155, + 0.009238882921636105, + 0.015440285205841064, + 0.013683326542377472, + 0.013698508031666279, + 0.012148956768214703, + 0.013551614247262478, + 0.004053313750773668, + 0.003576923394575715, + 0.005007022526115179, + 0.005192289128899574, + 0.004562684800475836, + 0.005507693625986576, + 0.006387877278029919, + 0.007432756945490837, + 0.0059943366795778275, + 0.013201714493334293, + 0.007056380156427622, + 0.011405745521187782, + 0.010217762552201748, + 0.009310591034591198, + 0.008955628611147404, + 0.009450805373489857, + 0.003548951121047139, + 0.0035796966403722763, + 0.0020541104022413492, + 0.0030748508870601654, + 0.0031743189319968224, + 0.005936434958130121, + 0.005137919448316097, + 0.008243954740464687, + 0.013082165271043777, + 0.009600411169230938, + 0.03529899939894676, + 0.023940667510032654, + 0.01529739424586296, + 0.2835824489593506, + 0.02617405727505684, + 0.25663870573043823, + 0.002816207706928253, + 0.0023768527898937464, + 0.0036536913830786943, + 0.0030695500317960978, + 0.00227616005577147, + 0.0032253589015454054, + 0.0020239672157913446, + 0.0019124036189168692, + 0.001763362786732614, + 0.0012108392547816038, + 0.0017108526080846786, + 0.001799847697839141, + 0.0016206977888941765, + 0.0019320441642776132, + 0.0014188564382493496, + 0.001775603974238038, + 0.002213485538959503, + 0.0017411077860742807, + 0.0016872418345883489, + 0.0025228792801499367, + 0.0017071946058422327, + 0.002069235546514392, + 0.0017098330426961184, + 0.002009087475016713, + 0.0018878525588661432, + 0.00256519066169858, + 0.002015498699620366, + 0.0027453608345240355, + 0.002264880808070302, + 0.0020538000389933586, + 0.0028831881936639547, + 0.0020053978078067303, + 0.0014534705551341176, + 0.0013247724855318666, + 0.0021080488804727793, + 0.0017054070485755801, + 0.0016208927845582366, + 0.0017892379546537995, + 0.000996821210719645, + 0.001977787120267749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07554557174444199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983570337295532 + ], + "opposite_logprob_tokens": [ + -0.5233570337295532 + ], + "logprob_reward": -0.8983570337295532, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983570337295532, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0345696322619915, + 0.05919617414474487, + 0.038723960518836975, + 0.049009330570697784, + 0.7010076642036438, + 0.03211214020848274, + 0.024312447756528854, + 0.013744959607720375, + 0.040691178292036057, + 0.006632548291236162 + ], + "credit_map": [ + 0.00244127050973475, + 0.003652112325653434, + 0.0035816251765936613, + 0.005899022798985243, + 0.0034719945397228003, + 0.004439171403646469, + 0.004928491543978453, + 0.006155940238386393, + 0.006350505631417036, + 0.005127592943608761, + 0.007741597015410662, + 0.007302208803594112, + 0.005087232682853937, + 0.00802003126591444, + 0.009561480954289436, + 0.010005521588027477, + 0.0039416407234966755, + 0.0047361277975142, + 0.003987726289778948, + 0.00528363510966301, + 0.004207827150821686, + 0.006753270514309406, + 0.004278605803847313, + 0.005535127129405737, + 0.00490372721105814, + 0.005435499828308821, + 0.004090932663530111, + 0.004742143675684929, + 0.005832838825881481, + 0.005092191975563765, + 0.007856221869587898, + 0.011055773124098778, + 0.013547451235353947, + 0.006062457337975502, + 0.03926940634846687, + 0.024266930297017097, + 0.013971584849059582, + 0.3077400028705597, + 0.027728958055377007, + 0.26842087507247925, + 0.0038910480216145515, + 0.0034365199971944094, + 0.0039489069022238255, + 0.003312895307317376, + 0.0027956105768680573, + 0.0053529562428593636, + 0.00504606356844306, + 0.0043281386606395245, + 0.0024870666675269604, + 0.0026335592847317457, + 0.0026755770668387413, + 0.0031104416120797396, + 0.0037941604387015104, + 0.0028978411573916674, + 0.0034016326535493135, + 0.0033121686428785324, + 0.0014386596158146858, + 0.001476642326451838, + 0.0016206834698095918, + 0.001477462355978787, + 0.001482422580011189, + 0.002021464752033353, + 0.0015185025986284018, + 0.002709121908992529, + 0.004079401958733797, + 0.0038492782041430473, + 0.005491793621331453, + 0.004171463660895824, + 0.004150563385337591, + 0.004766611848026514, + 0.008363505825400352, + 0.005818559788167477, + 0.0008291525300592184, + 0.00103947042953223, + 0.0008369371644221246, + 0.000924117339309305, + 0.0008602766320109367, + 0.0007089050486683846, + 0.0007206707377918065, + 0.0007130184094421566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.8983570337295532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003630477236583829 + ], + "opposite_logprob_tokens": [ + -5.628630638122559 + ], + "logprob_reward": -0.003630477236583829, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003630477236583829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0411163792014122, + 0.026021819561719894, + 0.07927596569061279, + 0.040298789739608765, + 0.7325460910797119, + 0.021415797993540764, + 0.015149583108723164, + 0.012887148186564445, + 0.02262817695736885, + 0.008660347200930119 + ], + "credit_map": [ + 0.003539057681336999, + 0.0037932726554572582, + 0.003986626397818327, + 0.007284797728061676, + 0.004985420498996973, + 0.007547543849796057, + 0.0048647234216332436, + 0.0051149302162230015, + 0.0027456493116915226, + 0.0035839031916111708, + 0.0029926977586001158, + 0.003473999910056591, + 0.0027874077204614878, + 0.0036170566454529762, + 0.0032515216153115034, + 0.0035695801489055157, + 0.00844620168209076, + 0.012936364859342575, + 0.009125957265496254, + 0.00852686632424593, + 0.00781785137951374, + 0.011646920815110207, + 0.00900315959006548, + 0.011772639118134975, + 0.004186047241091728, + 0.005425774492323399, + 0.004053723998367786, + 0.0056299325078725815, + 0.004532551392912865, + 0.004723985213786364, + 0.005826990585774183, + 0.005919782444834709, + 0.015061796642839909, + 0.006252089980989695, + 0.04101494327187538, + 0.02520003914833069, + 0.014244944788515568, + 0.3221043646335602, + 0.027284109964966774, + 0.2813837230205536, + 0.002558830427005887, + 0.0025482228957116604, + 0.0023824165109544992, + 0.001862151431851089, + 0.0025233773048967123, + 0.0036016078665852547, + 0.0029160918202251196, + 0.003023098222911358, + 0.0016800729790702462, + 0.0021508894860744476, + 0.0013983051758259535, + 0.001965608447790146, + 0.0019165672129020095, + 0.002463898155838251, + 0.0016039639012888074, + 0.0019702762365341187, + 0.0017646898049861193, + 0.001349970931187272, + 0.0014469173038378358, + 0.0013533291639760137, + 0.0017835225444287062, + 0.001979158027097583, + 0.0014844376128166914, + 0.0017251215176656842, + 0.0026576330419629812, + 0.003929168451577425, + 0.001986011629924178, + 0.0027271085418760777, + 0.0034262866247445345, + 0.00282038701698184, + 0.002739113522693515, + 0.0023424651008099318, + 0.0008398178033530712, + 0.0011017225915566087, + 0.0009720607195049524, + 0.0015701401280239224, + 0.0010469651315361261, + 0.0009259209036827087, + 0.0009357043309137225, + 0.0012680139625445008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.003630477236583829 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052617397159338 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.0052617397159338, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052617397159338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0452510342001915, + 0.029777292162179947, + 0.0369509756565094, + 0.031139731407165527, + 0.7478536367416382, + 0.02137029729783535, + 0.018812846392393112, + 0.02870030142366886, + 0.033096496015787125, + 0.007047458551824093 + ], + "credit_map": [ + 0.0048295739106833935, + 0.004099221900105476, + 0.0046871136873960495, + 0.008128797635436058, + 0.004019392188638449, + 0.008630864322185516, + 0.004600915126502514, + 0.006255153100937605, + 0.00319793657399714, + 0.004588515963405371, + 0.0031785096507519484, + 0.0033182110637426376, + 0.0037721809931099415, + 0.004517297726124525, + 0.0026349422987550497, + 0.004569698590785265, + 0.0022196031641215086, + 0.005814381409436464, + 0.0034001709427684546, + 0.006364370230585337, + 0.004016311839222908, + 0.004314327146857977, + 0.005644210614264011, + 0.005177601240575314, + 0.003432091325521469, + 0.0036011142656207085, + 0.003904967103153467, + 0.003862668527290225, + 0.003511367365717888, + 0.003412251127883792, + 0.0037511885166168213, + 0.0056640817783772945, + 0.015192515216767788, + 0.006535693071782589, + 0.04076318070292473, + 0.025567343458533287, + 0.01472522597759962, + 0.32866501808166504, + 0.02773268334567547, + 0.2886718809604645, + 0.0027551723178476095, + 0.0020107741001993418, + 0.0028709289617836475, + 0.0025577018968760967, + 0.00214496161788702, + 0.002845341106876731, + 0.0033689108677208424, + 0.0028165073599666357, + 0.0029109285678714514, + 0.0018094511469826102, + 0.001735609257593751, + 0.0019418211886659265, + 0.0018504966283217072, + 0.0028063312638550997, + 0.002442872617393732, + 0.0033153349068015814, + 0.005042023956775665, + 0.0018560306634753942, + 0.0039015160873532295, + 0.002281557535752654, + 0.002170328516513109, + 0.004419207107275724, + 0.003712983801960945, + 0.005316651426255703, + 0.005641436204314232, + 0.0033457784447818995, + 0.0039331731386482716, + 0.003693070961162448, + 0.003707482945173979, + 0.004602585919201374, + 0.0033415339421480894, + 0.004831436090171337, + 0.0007627036538906395, + 0.0008884213748387992, + 0.0008507092716172338, + 0.0009332153713330626, + 0.0010647284798324108, + 0.0007246750174090266, + 0.0007006634841673076, + 0.0011223417241126299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0052617397159338 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068257503211498 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.011068257503211498, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068257503211498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09017961472272873, + 0.03405057638883591, + 0.11579807847738266, + 0.0712200254201889, + 0.4896646738052368, + 0.0855649784207344, + 0.020434103906154633, + 0.041030675172805786, + 0.029126867651939392, + 0.022930411621928215 + ], + "credit_map": [ + 0.00844076368957758, + 0.01114638615399599, + 0.013061503879725933, + 0.01512109860777855, + 0.008978855796158314, + 0.010562537238001823, + 0.011162899434566498, + 0.011705579236149788, + 0.002652499359101057, + 0.005760501138865948, + 0.0031249509193003178, + 0.004649915266782045, + 0.0037690482567995787, + 0.005302810110151768, + 0.004050536546856165, + 0.004740315489470959, + 0.011437933892011642, + 0.019486336037516594, + 0.013355541974306107, + 0.013655771501362324, + 0.012533706612884998, + 0.014179150573909283, + 0.012875931337475777, + 0.018273720517754555, + 0.005634642206132412, + 0.0079946992918849, + 0.006538851652294397, + 0.007696087006479502, + 0.008720021694898605, + 0.008165251463651657, + 0.009888294152915478, + 0.0165821835398674, + 0.011995247565209866, + 0.0051968335174024105, + 0.030152935534715652, + 0.017694760113954544, + 0.01025043148547411, + 0.21354639530181885, + 0.018602123484015465, + 0.18222595751285553, + 0.01246979646384716, + 0.005895421840250492, + 0.014167947694659233, + 0.007246337365359068, + 0.00538300909101963, + 0.013730227947235107, + 0.01474087405949831, + 0.011931373737752438, + 0.002769921673461795, + 0.0020862610545009375, + 0.00281068729236722, + 0.0017657510470598936, + 0.0029433611780405045, + 0.0029870597645640373, + 0.0023838565684854984, + 0.0026872081216424704, + 0.006768934894353151, + 0.004731397610157728, + 0.004685227293521166, + 0.004865517374128103, + 0.004108721856027842, + 0.0055378819815814495, + 0.004194409120827913, + 0.006138591095805168, + 0.00485781766474247, + 0.004045951180160046, + 0.002659401623532176, + 0.004127570427954197, + 0.003842261154204607, + 0.0034897765144705772, + 0.002851995872333646, + 0.0032520948443561792, + 0.004479430615901947, + 0.002787180244922638, + 0.0031045028008520603, + 0.0032181874848902225, + 0.0020580925047397614, + 0.0021240722853690386, + 0.002549942582845688, + 0.002609005430713296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.011068257503211498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot framed in a vertical format, the camera tilts steadily downward from a slight downward angle, following a brass pocket watch, intricately engraved with vintage patterns, as it slowly slides down the surface of a tilted, deep burgundy silk handkerchief. The aged brass case, softly lit from the side, gleams with a warm, deep metallic luster. The interplay of light and shadow lends the surface a vivid three-dimensional quality, while the slight wear on the edges betrays the wear of time. The delicate texture of the silk handkerchief is clearly visible, its satiny sheen shifting and changing with the movement of the watch. The watch's descent is not sudden, but rather steady and deliberate, its weight indenting the soft silk with a shallow, shifting indentation. The camera precisely captures every subtle lag between metal and fabric, almost perceiving the silent resistance. The entire picture is shrouded in a tranquil and nostalgic atmosphere. The high level of detail captured creates a sharp contrast between the cold hardness of brass and the soft warmth of silk, creating an elegant meditation on the passage of time and the beauty of material things.", + "seed": 2479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_9229_seed_2479_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.23507350385189058, + "vlm_logprob": -0.23507350385189058, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.23507350385189058, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09055470675230026 + ], + "opposite_logprob_tokens": [ + -3.090554714202881 + ], + "logprob_reward": -0.09055470675230026, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09055470675230026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46764931082725525, + 0.10357780754566193, + 0.12679491937160492, + 0.08398905396461487, + 0.028448544442653656, + 0.023208556696772575, + 0.032819315791130066, + 0.04594392701983452, + 0.03510861098766327, + 0.052459921687841415 + ], + "credit_map": [ + 0.028691967949271202, + 0.03634582459926605, + 0.07578723132610321, + 0.019109271466732025, + 0.02418414130806923, + 0.11107910424470901, + 0.02710300125181675, + 0.14534878730773926, + 0.00608134176582098, + 0.007145756855607033, + 0.026673171669244766, + 0.009553779847919941, + 0.011638275347650051, + 0.006400858983397484, + 0.01290802750736475, + 0.02317660301923752, + 0.007213944103568792, + 0.005991836078464985, + 0.036962468177080154, + 0.009256965480744839, + 0.030621709302067757, + 0.0064025926403701305, + 0.01658708043396473, + 0.01375833060592413, + 0.007947173900902271, + 0.007542919833213091, + 0.012303197756409645, + 0.008029970340430737, + 0.013955606147646904, + 0.013020788319408894, + 0.009299122728407383, + 0.011890278197824955, + 0.002792708110064268, + 0.002227196004241705, + 0.0035359193570911884, + 0.0030154851265251637, + 0.002419414697214961, + 0.005300892051309347, + 0.0046685971319675446, + 0.004488333594053984, + 0.0017463277326896787, + 0.002108086831867695, + 0.005556533113121986, + 0.001904491800814867, + 0.0017120047705247998, + 0.0036449942272156477, + 0.002844708040356636, + 0.0036914132069796324, + 0.0024072257801890373, + 0.0035708434879779816, + 0.00380852073431015, + 0.004573955200612545, + 0.0038966992869973183, + 0.005327478516846895, + 0.004499363712966442, + 0.004735234193503857, + 0.0033913596998900175, + 0.006564532406628132, + 0.005388319492340088, + 0.005207002628594637, + 0.0054961335845291615, + 0.005972369574010372, + 0.00854500476270914, + 0.00537920743227005, + 0.003364800475537777, + 0.003523872932419181, + 0.004294842015951872, + 0.0033545400947332382, + 0.004420355893671513, + 0.0050073848105967045, + 0.0059069739654660225, + 0.005235843826085329, + 0.005617646500468254, + 0.004380353726446629, + 0.0070622763596475124, + 0.005851031746715307, + 0.006280804984271526, + 0.007943771779537201, + 0.007815280929207802, + 0.007508762180805206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.09055470675230026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539228916168213 + ], + "opposite_logprob_tokens": [ + -0.4289228618144989 + ], + "logprob_reward": -1.0539228916168213, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539228916168213, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5443374514579773, + 0.0355953574180603, + 0.058509521186351776, + 0.04004964232444763, + 0.02459854446351528, + 0.030248189345002174, + 0.029491208493709564, + 0.09712276607751846, + 0.08927584439516068, + 0.05077138543128967 + ], + "credit_map": [ + 0.024326030164957047, + 0.01710703782737255, + 0.22220993041992188, + 0.022545799612998962, + 0.02041582204401493, + 0.16910983622074127, + 0.01335948146879673, + 0.05526358261704445, + 0.0038060196675360203, + 0.0036824981216341257, + 0.0027752448804676533, + 0.003782392479479313, + 0.00658322311937809, + 0.003766539739444852, + 0.0059647089801728725, + 0.005234734620898962, + 0.0034642864484339952, + 0.0044069536961615086, + 0.015966646373271942, + 0.004378222860395908, + 0.011047466658055782, + 0.005781979765743017, + 0.006274302490055561, + 0.0071896761655807495, + 0.004140791017562151, + 0.0037717181257903576, + 0.010226378217339516, + 0.0032312702387571335, + 0.0034247238654643297, + 0.007272947579622269, + 0.0038113410118967295, + 0.004170479252934456, + 0.0023905555717647076, + 0.0022442599292844534, + 0.0041004810482263565, + 0.0022192811593413353, + 0.0015406524762511253, + 0.002902835374698043, + 0.0027463785372674465, + 0.006454103160649538, + 0.002006813883781433, + 0.0019673488568514585, + 0.007652120664715767, + 0.002647326560690999, + 0.0016627553850412369, + 0.005094926338642836, + 0.002744968282058835, + 0.006471933797001839, + 0.0016442470951005816, + 0.0028342637233436108, + 0.002955289091914892, + 0.004886386916041374, + 0.004763598553836346, + 0.003440392669290304, + 0.005089164711534977, + 0.003877870040014386, + 0.00503937155008316, + 0.011819642968475819, + 0.011060434393584728, + 0.013466454111039639, + 0.014306202530860901, + 0.014434034936130047, + 0.015347116626799107, + 0.011649521999061108, + 0.006741814315319061, + 0.008213277906179428, + 0.008338513784110546, + 0.010819725692272186, + 0.013221180997788906, + 0.015638215467333794, + 0.014503590762615204, + 0.011799542233347893, + 0.004111803602427244, + 0.004972789902240038, + 0.00530760595574975, + 0.005402152426540852, + 0.007045669015496969, + 0.005840861704200506, + 0.010675466619431973, + 0.0074150450527668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -1.0539228916168213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750516593456268 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750516593456268, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750516593456268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.464603066444397, + 0.08736857771873474, + 0.14649146795272827, + 0.05877941846847534, + 0.04574774578213692, + 0.03956193849444389, + 0.025380335748195648, + 0.019283462315797806, + 0.02205279842019081, + 0.09073124080896378 + ], + "credit_map": [ + 0.050058215856552124, + 0.029313035309314728, + 0.12536314129829407, + 0.02779441513121128, + 0.02150542289018631, + 0.12798985838890076, + 0.04583050310611725, + 0.03674847260117531, + 0.006566727999597788, + 0.009765753522515297, + 0.013276178389787674, + 0.009060031734406948, + 0.021642398089170456, + 0.007643119432032108, + 0.011568507179617882, + 0.007845867425203323, + 0.008641852997243404, + 0.00994348619133234, + 0.04325861111283302, + 0.013526531867682934, + 0.031031787395477295, + 0.008115625940263271, + 0.018883638083934784, + 0.013089930638670921, + 0.003383193165063858, + 0.005240251775830984, + 0.015285985544323921, + 0.004316018894314766, + 0.006552666425704956, + 0.006499938201159239, + 0.00492828618735075, + 0.012573080137372017, + 0.0037330116610974073, + 0.003483403939753771, + 0.005947626195847988, + 0.0046741049736738205, + 0.002781978342682123, + 0.0048048715107142925, + 0.006841257214546204, + 0.01348149310797453, + 0.002840087516233325, + 0.002744411351159215, + 0.009209444746375084, + 0.003667366923764348, + 0.0020593865774571896, + 0.007067681755870581, + 0.0035411573480814695, + 0.008432400412857533, + 0.002052590250968933, + 0.0020995456725358963, + 0.003304202575236559, + 0.0032084977719932795, + 0.003162666689604521, + 0.0033943946473300457, + 0.004845866933465004, + 0.0033125721383839846, + 0.0016334251267835498, + 0.0023861245717853308, + 0.001503950683400035, + 0.0017340099439024925, + 0.0019686268642544746, + 0.00368691049516201, + 0.0037196313496679068, + 0.0026507831644266844, + 0.002124330960214138, + 0.0017352666473016143, + 0.0032559323590248823, + 0.0019831396639347076, + 0.0031085654627531767, + 0.0024032588116824627, + 0.0040892427787184715, + 0.003353061620146036, + 0.007028764113783836, + 0.009473077021539211, + 0.009110964834690094, + 0.009430835954844952, + 0.014296237379312515, + 0.010066801682114601, + 0.018911415711045265, + 0.012413140386343002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006750516593456268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959006026387215 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005959006026387215, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959006026387215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8170220851898193, + 0.03515511378645897, + 0.02525765262544155, + 0.025439592078328133, + 0.011811194941401482, + 0.010172679089009762, + 0.03030688688158989, + 0.024981558322906494, + 0.008588450029492378, + 0.011264816857874393 + ], + "credit_map": [ + 0.03939240798354149, + 0.025552667677402496, + 0.33580926060676575, + 0.02978036366403103, + 0.03168535977602005, + 0.263643741607666, + 0.013519168831408024, + 0.07763903588056564, + 0.0025021606124937534, + 0.0027536002453416586, + 0.008580129593610764, + 0.003162575652822852, + 0.006087951827794313, + 0.0023705156054347754, + 0.005031246691942215, + 0.00466693053022027, + 0.001661328598856926, + 0.0018300270894542336, + 0.006166927516460419, + 0.0021373291965574026, + 0.005327839870005846, + 0.0025597712956368923, + 0.0026690419763326645, + 0.0029053848702460527, + 0.0018164505017921329, + 0.001792028546333313, + 0.0077978516928851604, + 0.0018967714859172702, + 0.0024684967938810587, + 0.0034139526542276144, + 0.0025632192846387625, + 0.003690818091854453, + 0.001127227209508419, + 0.00113514659460634, + 0.0015733070904389024, + 0.001420017797499895, + 0.000927445653360337, + 0.0016453000716865063, + 0.0020951279439032078, + 0.0018876211252063513, + 0.0006897160201333463, + 0.000932839815504849, + 0.0018974434351548553, + 0.0010815588757395744, + 0.0006987502565607429, + 0.0013975348556414247, + 0.0013776886044070125, + 0.0020971461199223995, + 0.002394316019490361, + 0.0024180009495466948, + 0.0043646604754030704, + 0.003658383386209607, + 0.002753550885245204, + 0.004910366144031286, + 0.005188295617699623, + 0.004619311541318893, + 0.0021654756274074316, + 0.0032650113571435213, + 0.0021656956523656845, + 0.0033896651584655046, + 0.0021299647632986307, + 0.004094626288861036, + 0.0050986758433282375, + 0.0026724408380687237, + 0.0010824507335200906, + 0.0009484465699642897, + 0.0007010396220721304, + 0.000968251028098166, + 0.0011736323358491063, + 0.0013062201905995607, + 0.0016261714044958353, + 0.0007822370389476418, + 0.0015347517328336835, + 0.0013160710223019123, + 0.001246331143192947, + 0.0009062408353202045, + 0.0013648908352479339, + 0.0013035709271207452, + 0.0022323622833937407, + 0.0013605966232717037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.005959006026387215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4233298897743225, + 0.1186717227101326, + 0.05670519918203354, + 0.05513644218444824, + 0.11484245955944061, + 0.05700954794883728, + 0.033399712294340134, + 0.0252678319811821, + 0.028727274388074875, + 0.08690997958183289 + ], + "credit_map": [ + 0.04495954513549805, + 0.029249748215079308, + 0.10157250612974167, + 0.029142355546355247, + 0.021158037707209587, + 0.11800627410411835, + 0.04328525438904762, + 0.03595616668462753, + 0.006680566351860762, + 0.012266253121197224, + 0.026600131765007973, + 0.009540223516523838, + 0.031696487218141556, + 0.008522441610693932, + 0.012880947440862656, + 0.010484670288860798, + 0.0035539912059903145, + 0.004124373197555542, + 0.014732462354004383, + 0.0043425546027719975, + 0.009981024079024792, + 0.006598408333957195, + 0.006224335171282291, + 0.007148054428398609, + 0.003241563681513071, + 0.004614317789673805, + 0.015915922820568085, + 0.003536535892635584, + 0.0047215684317052364, + 0.005776991602033377, + 0.0042192074470222, + 0.013110334984958172, + 0.009010213427245617, + 0.009403727948665619, + 0.019745536148548126, + 0.008819963783025742, + 0.006242471281439066, + 0.014163604006171227, + 0.009214590303599834, + 0.03824235126376152, + 0.004425260704010725, + 0.0038385302759706974, + 0.013824394904077053, + 0.004812446888536215, + 0.0026775426231324673, + 0.010298969224095345, + 0.004548841621726751, + 0.012583563104271889, + 0.002447721315547824, + 0.003424476832151413, + 0.006100446451455355, + 0.003425637958571315, + 0.003139750100672245, + 0.004432563669979572, + 0.003970912657678127, + 0.006458205170929432, + 0.0017851062584668398, + 0.002399856224656105, + 0.002310914918780327, + 0.0025406640488654375, + 0.0022860541939735413, + 0.00470846937969327, + 0.006037585902959108, + 0.0031991812866181135, + 0.0022554986644536257, + 0.002709310967475176, + 0.0036336868070065975, + 0.0024818102829158306, + 0.004588892683386803, + 0.004029578994959593, + 0.00537090701982379, + 0.0036575894337147474, + 0.006539760157465935, + 0.008165379986166954, + 0.009153061546385288, + 0.009005023166537285, + 0.01209041103720665, + 0.011374378576874733, + 0.017294157296419144, + 0.013287815265357494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot framed in a vertical format, the camera tilts steadily downward from a slight downward angle, following a brass pocket watch, intricately engraved with vintage patterns, as it slowly slides down the surface of a tilted, deep burgundy silk handkerchief. The aged brass case, softly lit from the side, gleams with a warm, deep metallic luster. The interplay of light and shadow lends the surface a vivid three-dimensional quality, while the slight wear on the edges betrays the wear of time. The delicate texture of the silk handkerchief is clearly visible, its satiny sheen shifting and changing with the movement of the watch. The watch's descent is not sudden, but rather steady and deliberate, its weight indenting the soft silk with a shallow, shifting indentation. The camera precisely captures every subtle lag between metal and fabric, almost perceiving the silent resistance. The entire picture is shrouded in a tranquil and nostalgic atmosphere. The high level of detail captured creates a sharp contrast between the cold hardness of brass and the soft warmth of silk, creating an elegant meditation on the passage of time and the beauty of material things.", + "seed": 2480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_9229_seed_2480_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15566276013851166, + "vlm_logprob": -0.15566276013851166, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.15566276013851166, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06697629392147064 + ], + "opposite_logprob_tokens": [ + -3.441976308822632 + ], + "logprob_reward": -0.06697629392147064, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06697629392147064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19392625987529755, + 0.09443403780460358, + 0.0647285133600235, + 0.0538535937666893, + 0.14471158385276794, + 0.19748179614543915, + 0.05726305767893791, + 0.06625653058290482, + 0.06687648594379425, + 0.06046806275844574 + ], + "credit_map": [ + 0.025279488414525986, + 0.04926738142967224, + 0.022365916520357132, + 0.019715873524546623, + 0.01957240328192711, + 0.01973937451839447, + 0.020525241270661354, + 0.017460575327277184, + 0.010167140513658524, + 0.009811192750930786, + 0.014088255353271961, + 0.006575314328074455, + 0.01614684984087944, + 0.018428098410367966, + 0.006761875469237566, + 0.012455317191779613, + 0.006867533084005117, + 0.005704840645194054, + 0.006847829092293978, + 0.006736615672707558, + 0.01206041406840086, + 0.008282245136797428, + 0.007094541564583778, + 0.011134489439427853, + 0.006217983551323414, + 0.005306667648255825, + 0.006578607019037008, + 0.008161160163581371, + 0.005755616817623377, + 0.006940394174307585, + 0.007518379483371973, + 0.007374784909188747, + 0.011166413314640522, + 0.006674342788755894, + 0.013969149440526962, + 0.007494122721254826, + 0.015073818154633045, + 0.051041848957538605, + 0.01009906642138958, + 0.029192814603447914, + 0.06332939863204956, + 0.007526356261223555, + 0.017635978758335114, + 0.005811307579278946, + 0.017396414652466774, + 0.06890297681093216, + 0.005680848378688097, + 0.011198491789400578, + 0.008870198391377926, + 0.0075827655382454395, + 0.005086611025035381, + 0.008545970544219017, + 0.009336964227259159, + 0.006756120827049017, + 0.006648415699601173, + 0.004436012357473373, + 0.00819565262645483, + 0.005968999583274126, + 0.007389482576400042, + 0.005903787445276976, + 0.005405009258538485, + 0.01526580099016428, + 0.004200884606689215, + 0.013926916755735874, + 0.00666993809863925, + 0.007966696284711361, + 0.015163903124630451, + 0.00799254234880209, + 0.00568421883508563, + 0.009979031048715115, + 0.007169673219323158, + 0.00625048391520977, + 0.007027598563581705, + 0.00820721872150898, + 0.0076652816496789455, + 0.010475619696080685, + 0.006095028482377529, + 0.007311704568564892, + 0.007757805287837982, + 0.0059278071857988834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.06697629392147064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933982968330383 + ], + "opposite_logprob_tokens": [ + -0.6933982968330383 + ], + "logprob_reward": -0.6933982968330383, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933982968330383, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09243148565292358, + 0.03829272463917732, + 0.045720186084508896, + 0.05050725117325783, + 0.14266841113567352, + 0.09534724801778793, + 0.06121968477964401, + 0.18499374389648438, + 0.1637508124113083, + 0.12506839632987976 + ], + "credit_map": [ + 0.011744007468223572, + 0.025427022948861122, + 0.0070261904038488865, + 0.011464833281934261, + 0.009592864662408829, + 0.009621936827898026, + 0.009293402545154095, + 0.008261230774223804, + 0.0032549984753131866, + 0.004166011232882738, + 0.0047273095697164536, + 0.004260613117367029, + 0.005875702947378159, + 0.005890770349651575, + 0.004740914795547724, + 0.005376406013965607, + 0.005452386569231749, + 0.0036830001045018435, + 0.004517335444688797, + 0.004659105557948351, + 0.008112526498734951, + 0.005313232075423002, + 0.007043005898594856, + 0.006939596962183714, + 0.006698362994939089, + 0.005107394885271788, + 0.006360008381307125, + 0.004675097763538361, + 0.006451943423599005, + 0.008426683954894543, + 0.005643076729029417, + 0.007144681643694639, + 0.010449171997606754, + 0.0033797970972955227, + 0.016825931146740913, + 0.003736282465979457, + 0.015823930501937866, + 0.060943517833948135, + 0.007284493651241064, + 0.024225281551480293, + 0.012506283819675446, + 0.010036799125373363, + 0.013663498684763908, + 0.015766119584441185, + 0.007158919703215361, + 0.010765948332846165, + 0.013226104900240898, + 0.012223574332892895, + 0.010282323695719242, + 0.005612982902675867, + 0.006127022206783295, + 0.004975432064384222, + 0.012732005678117275, + 0.011063752695918083, + 0.005706296302378178, + 0.004719871561974287, + 0.012589851394295692, + 0.006570847239345312, + 0.021867401897907257, + 0.006103353574872017, + 0.01998675987124443, + 0.07501323521137238, + 0.00974617712199688, + 0.03311610966920853, + 0.01693856157362461, + 0.01283229235559702, + 0.052940733730793, + 0.012502863071858883, + 0.012674876488745213, + 0.019704043865203857, + 0.012388785369694233, + 0.023768646642565727, + 0.011700383387506008, + 0.017667217180132866, + 0.02116892859339714, + 0.017425794154405594, + 0.013893621042370796, + 0.018957041203975677, + 0.012514917179942131, + 0.01174050010740757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.6933982968330383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022074636071920395 + ], + "opposite_logprob_tokens": [ + -6.127207279205322 + ], + "logprob_reward": -0.0022074636071920395, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022074636071920395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13787342607975006, + 0.06357526034116745, + 0.04670402407646179, + 0.07576567679643631, + 0.09743019193410873, + 0.36585095524787903, + 0.025549858808517456, + 0.04398304224014282, + 0.08821582049131393, + 0.05505171790719032 + ], + "credit_map": [ + 0.016233058646321297, + 0.03361032158136368, + 0.013421202078461647, + 0.015079031698405743, + 0.01913798600435257, + 0.0174941997975111, + 0.010489960201084614, + 0.012407680973410606, + 0.004984069615602493, + 0.009066592901945114, + 0.006081575062125921, + 0.007595451548695564, + 0.007126546464860439, + 0.009944017045199871, + 0.006447576452046633, + 0.01232944056391716, + 0.004319210071116686, + 0.003493750235065818, + 0.0069269947707653046, + 0.00513088796287775, + 0.007919613271951675, + 0.006560747977346182, + 0.0047907354310154915, + 0.007562087383121252, + 0.004979624878615141, + 0.004213753622025251, + 0.006942682899534702, + 0.006104881409555674, + 0.0068953814916312695, + 0.024394506588578224, + 0.006101242266595364, + 0.016133612021803856, + 0.006389100104570389, + 0.005723890848457813, + 0.013295563869178295, + 0.004797426983714104, + 0.008923886343836784, + 0.035895369946956635, + 0.006339642684906721, + 0.01606532372534275, + 0.11386922746896744, + 0.014417847618460655, + 0.03473436087369919, + 0.01046687550842762, + 0.03020334802567959, + 0.13357871770858765, + 0.009871009737253189, + 0.018709592521190643, + 0.0032468880526721478, + 0.003297625109553337, + 0.0025127006229013205, + 0.004675730597227812, + 0.0024033510126173496, + 0.003951725549995899, + 0.003214664990082383, + 0.002247175667434931, + 0.006299749482423067, + 0.005058840848505497, + 0.005643226206302643, + 0.004564499482512474, + 0.004323533736169338, + 0.008978892117738724, + 0.005084719974547625, + 0.004029585514217615, + 0.00973794236779213, + 0.010319286026060581, + 0.018968017771840096, + 0.010829919017851353, + 0.006525914650410414, + 0.014082014560699463, + 0.007991054095327854, + 0.009761671535670757, + 0.007264836225658655, + 0.00686816917732358, + 0.006094347685575485, + 0.007174348924309015, + 0.004195226822048426, + 0.009469778276979923, + 0.007681385613977909, + 0.006303629372268915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0022074636071920395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031987475231289864 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031987475231289864, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031987475231289864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0941130667924881, + 0.07700216770172119, + 0.05522488057613373, + 0.13949111104011536, + 0.02978595346212387, + 0.38096389174461365, + 0.01812928356230259, + 0.13129548728466034, + 0.022884640842676163, + 0.051109567284584045 + ], + "credit_map": [ + 0.009486564435064793, + 0.025433214381337166, + 0.00846811756491661, + 0.006192192900925875, + 0.012046598829329014, + 0.014222496189177036, + 0.009314408525824547, + 0.008949480019509792, + 0.009489597752690315, + 0.009735054336488247, + 0.012389740906655788, + 0.007668158505111933, + 0.008627320639789104, + 0.008877928368747234, + 0.00726437009871006, + 0.012950003147125244, + 0.005438921973109245, + 0.004214366897940636, + 0.008090484887361526, + 0.008683779276907444, + 0.0077676186338067055, + 0.0064767589792609215, + 0.00700062420219183, + 0.007552322465926409, + 0.012569145299494267, + 0.008180397562682629, + 0.01028321124613285, + 0.01047535240650177, + 0.011237427592277527, + 0.05125538632273674, + 0.010155546478927135, + 0.02533464878797531, + 0.0032521202228963375, + 0.0028538156766444445, + 0.0037551585119217634, + 0.004066255409270525, + 0.002737040864303708, + 0.0048501756973564625, + 0.004393371287733316, + 0.0038780136965215206, + 0.10634420067071915, + 0.012401857413351536, + 0.039910364896059036, + 0.011503370478749275, + 0.03261348605155945, + 0.14694644510746002, + 0.009698108769953251, + 0.021546069532632828, + 0.0018434901721775532, + 0.002276771469041705, + 0.0016530503053218126, + 0.0030460578855127096, + 0.002031286945566535, + 0.0025455474387854338, + 0.002335372380912304, + 0.002397708361968398, + 0.009839832782745361, + 0.003741129767149687, + 0.018247466534376144, + 0.0039605735801160336, + 0.01469025295227766, + 0.05584380775690079, + 0.004755440633744001, + 0.020216990262269974, + 0.0022188187576830387, + 0.0031186218839138746, + 0.004484135657548904, + 0.0025926977396011353, + 0.0017577992985025048, + 0.003448060480877757, + 0.0029326246585696936, + 0.002331883180886507, + 0.005107223521918058, + 0.006587625946849585, + 0.00699276477098465, + 0.00767430430278182, + 0.006241337396204472, + 0.008025990799069405, + 0.0056338561698794365, + 0.004846467170864344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.0031987475231289864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253299880772829 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253299880772829, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253299880772829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149942085146904, + 0.07100319862365723, + 0.032305486500263214, + 0.06870701164007187, + 0.0638091191649437, + 0.3259566128253937, + 0.026886120438575745, + 0.08600753545761108, + 0.07104366272687912, + 0.10433913767337799 + ], + "credit_map": [ + 0.016504749655723572, + 0.03099171631038189, + 0.016410019248723984, + 0.01169358566403389, + 0.02523762360215187, + 0.020989706739783287, + 0.01101283635944128, + 0.0171018298715353, + 0.0075662145391106606, + 0.009536507539451122, + 0.00918160006403923, + 0.0072068097069859505, + 0.007956048473715782, + 0.013069657608866692, + 0.005328991450369358, + 0.011157370172441006, + 0.0033104252070188522, + 0.003023169469088316, + 0.00380046502687037, + 0.0033984053879976273, + 0.0054718912579119205, + 0.004449681844562292, + 0.004473534878343344, + 0.00437791645526886, + 0.008005587384104729, + 0.004034390673041344, + 0.008768744766712189, + 0.00407163193449378, + 0.008466149680316448, + 0.02149319089949131, + 0.004689393099397421, + 0.009177922271192074, + 0.015043488703668118, + 0.006820627488195896, + 0.008535338565707207, + 0.006713021546602249, + 0.00446016900241375, + 0.007392610888928175, + 0.006171681918203831, + 0.008672178722918034, + 0.09619440883398056, + 0.01926039531826973, + 0.03386083245277405, + 0.015177574940025806, + 0.026378002017736435, + 0.09459643065929413, + 0.018046388402581215, + 0.02244258113205433, + 0.003173964563757181, + 0.002602978842332959, + 0.0024010927882045507, + 0.002672287868335843, + 0.0052701798267662525, + 0.00514946086332202, + 0.003035247093066573, + 0.0025809097569435835, + 0.0046509504318237305, + 0.00392576027661562, + 0.010594503954052925, + 0.0032616995740681887, + 0.009534735232591629, + 0.03459435701370239, + 0.003925389144569635, + 0.015520134940743446, + 0.007287428714334965, + 0.007496626581996679, + 0.015835851430892944, + 0.007595251780003309, + 0.005218467675149441, + 0.01019800640642643, + 0.009523729793727398, + 0.007888306863605976, + 0.014530589804053307, + 0.0142434136942029, + 0.012140153907239437, + 0.012232566252350807, + 0.009099218994379044, + 0.01698337122797966, + 0.013947696425020695, + 0.011162133887410164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.01253299880772829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot framed in a vertical format, the camera tilts steadily downward from a slight downward angle, following a brass pocket watch, intricately engraved with vintage patterns, as it slowly slides down the surface of a tilted, deep burgundy silk handkerchief. The aged brass case, softly lit from the side, gleams with a warm, deep metallic luster. The interplay of light and shadow lends the surface a vivid three-dimensional quality, while the slight wear on the edges betrays the wear of time. The delicate texture of the silk handkerchief is clearly visible, its satiny sheen shifting and changing with the movement of the watch. The watch's descent is not sudden, but rather steady and deliberate, its weight indenting the soft silk with a shallow, shifting indentation. The camera precisely captures every subtle lag between metal and fabric, almost perceiving the silent resistance. The entire picture is shrouded in a tranquil and nostalgic atmosphere. The high level of detail captured creates a sharp contrast between the cold hardness of brass and the soft warmth of silk, creating an elegant meditation on the passage of time and the beauty of material things.", + "seed": 2481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_8/index_9229_seed_2481_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17083415689412504, + "vlm_logprob": -0.17083415689412504, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.17083415689412504, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07554513216018677 + ], + "opposite_logprob_tokens": [ + -3.325545072555542 + ], + "logprob_reward": -0.07554513216018677, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.07554513216018677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22619786858558655, + 0.11480530351400375, + 0.0976639986038208, + 0.07802372425794601, + 0.07595551759004593, + 0.09659828990697861, + 0.11046655476093292, + 0.06642709672451019, + 0.08793502300977707, + 0.04592665284872055 + ], + "credit_map": [ + 0.020458944141864777, + 0.03431272506713867, + 0.032104432582855225, + 0.02420896477997303, + 0.023366397246718407, + 0.03501356765627861, + 0.028223419561982155, + 0.028509438037872314, + 0.009941193275153637, + 0.013047246262431145, + 0.013857086189091206, + 0.01643993891775608, + 0.011905546300113201, + 0.015322316437959671, + 0.017672350630164146, + 0.01661962829530239, + 0.009312028996646404, + 0.009652338922023773, + 0.01443540956825018, + 0.0181904099881649, + 0.010415749624371529, + 0.008997578173875809, + 0.01492056343704462, + 0.01173991896212101, + 0.009655985049903393, + 0.013786310330033302, + 0.010796360671520233, + 0.012812238186597824, + 0.007783814799040556, + 0.00496832886710763, + 0.010603006929159164, + 0.007617676630616188, + 0.006332182791084051, + 0.010818376205861568, + 0.011472196318209171, + 0.010166587308049202, + 0.007786605041474104, + 0.0066428519785404205, + 0.011371145956218243, + 0.011365571059286594, + 0.007138648070394993, + 0.011869074776768684, + 0.02535911276936531, + 0.012142420746386051, + 0.006403195671737194, + 0.004869830794632435, + 0.012292093597352505, + 0.01652390882372856, + 0.008237886242568493, + 0.005223082844167948, + 0.006600355729460716, + 0.007103570271283388, + 0.005938273388892412, + 0.0054941377602517605, + 0.03640769049525261, + 0.035461556166410446, + 0.007390409242361784, + 0.0037361159920692444, + 0.019919762387871742, + 0.005025653168559074, + 0.003311938839033246, + 0.0038924282416701317, + 0.014111345633864403, + 0.009039443917572498, + 0.009775189682841301, + 0.005712962243705988, + 0.011856521479785442, + 0.012377209030091763, + 0.004567602649331093, + 0.007851392962038517, + 0.015800604596734047, + 0.01999354548752308, + 0.004495040979236364, + 0.0022946004755795, + 0.008709670975804329, + 0.004105446394532919, + 0.002905565779656172, + 0.003465586807578802, + 0.006751656997948885, + 0.013199085369706154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.07554513216018677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578448057174683 + ], + "opposite_logprob_tokens": [ + -0.6328448057174683 + ], + "logprob_reward": -0.7578448057174683, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578448057174683, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1275463104248047, + 0.07754228264093399, + 0.049182552844285965, + 0.11187461763620377, + 0.07846327126026154, + 0.12928488850593567, + 0.2551603615283966, + 0.06383863836526871, + 0.09845422208309174, + 0.0086528230458498 + ], + "credit_map": [ + 0.008164898492395878, + 0.01294192485511303, + 0.012185820378363132, + 0.011462315917015076, + 0.010827841237187386, + 0.022701481357216835, + 0.012169113382697105, + 0.037092920392751694, + 0.005703771952539682, + 0.007903095334768295, + 0.012448660098016262, + 0.012257478199899197, + 0.008231665007770061, + 0.008637992665171623, + 0.012855993583798409, + 0.009503625333309174, + 0.00557864410802722, + 0.0048319618217647076, + 0.007114428095519543, + 0.005257996264845133, + 0.005303988233208656, + 0.005065088160336018, + 0.005566728767007589, + 0.010463714599609375, + 0.011147669516503811, + 0.022391077131032944, + 0.01460078451782465, + 0.017294391989707947, + 0.011703116819262505, + 0.00946980994194746, + 0.0131372744217515, + 0.012130496092140675, + 0.006828532088547945, + 0.003437404753640294, + 0.0230899415910244, + 0.007493368349969387, + 0.004072748590260744, + 0.009592097252607346, + 0.01294912863522768, + 0.011000052094459534, + 0.008967539295554161, + 0.007292207796126604, + 0.041904859244823456, + 0.014520149677991867, + 0.0065089259296655655, + 0.010283348150551319, + 0.014132698066532612, + 0.025675153359770775, + 0.016427043825387955, + 0.0076011004857718945, + 0.06063881143927574, + 0.015206214971840382, + 0.010456436313688755, + 0.010814759880304337, + 0.09391710162162781, + 0.04009890556335449, + 0.0044628167524933815, + 0.005247524473816156, + 0.012156907469034195, + 0.005280718207359314, + 0.004207570105791092, + 0.0037156962789595127, + 0.018432367593050003, + 0.010335037484765053, + 0.00932937953621149, + 0.011583862826228142, + 0.013453255407512188, + 0.012483115307986736, + 0.00927821733057499, + 0.008313783444464207, + 0.018010416999459267, + 0.016002198681235313, + 0.0010663371067494154, + 0.0006094829877838492, + 0.0014641019515693188, + 0.0007779252482578158, + 0.000650116999167949, + 0.0007817625300958753, + 0.0011592464288696647, + 0.002143849153071642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.7578448057174683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019507441902533174 + ], + "opposite_logprob_tokens": [ + -6.251950740814209 + ], + "logprob_reward": -0.0019507441902533174, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019507441902533174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3489104211330414, + 0.15906719863414764, + 0.052721913903951645, + 0.05795009061694145, + 0.042615313082933426, + 0.08003614842891693, + 0.14307735860347748, + 0.0556449294090271, + 0.030360998585820198, + 0.02961570769548416 + ], + "credit_map": [ + 0.02822951041162014, + 0.056241076439619064, + 0.03830982744693756, + 0.026590976864099503, + 0.028690282255411148, + 0.0648040696978569, + 0.030889110639691353, + 0.07515552639961243, + 0.010167602449655533, + 0.012498346157371998, + 0.04528932273387909, + 0.014926185831427574, + 0.011484342627227306, + 0.015531250275671482, + 0.013538967818021774, + 0.035631172358989716, + 0.004830287769436836, + 0.004667623434215784, + 0.00883366260677576, + 0.006460633594542742, + 0.005549820140004158, + 0.006365182809531689, + 0.00591242453083396, + 0.01010227669030428, + 0.006147116422653198, + 0.007422813680022955, + 0.007683640345931053, + 0.005856527481228113, + 0.00837805587798357, + 0.006075280252844095, + 0.008909445255994797, + 0.00747720617800951, + 0.004481464624404907, + 0.004823296330869198, + 0.008752460591495037, + 0.004974861163645983, + 0.0024170910473912954, + 0.005138223525136709, + 0.00400301581248641, + 0.008024901151657104, + 0.006783557590097189, + 0.004129692446440458, + 0.025082595646381378, + 0.00706271780654788, + 0.003862569108605385, + 0.006161120720207691, + 0.006251255050301552, + 0.02070263773202896, + 0.00968047883361578, + 0.007572996895760298, + 0.04252489656209946, + 0.010785585269331932, + 0.0076678055338561535, + 0.007779144216328859, + 0.01238952949643135, + 0.04467691481113434, + 0.005892645101994276, + 0.0034046745859086514, + 0.0044228327460587025, + 0.00483576999977231, + 0.0030583455227315426, + 0.003615943482145667, + 0.016495512798428535, + 0.013919200748205185, + 0.0039687505923211575, + 0.0034060070756822824, + 0.004071944393217564, + 0.0037222523242235184, + 0.002947611501440406, + 0.0025394202675670385, + 0.004807323683053255, + 0.0048976861871778965, + 0.0035978322848677635, + 0.0022047534584999084, + 0.004798647947609425, + 0.0038444388192147017, + 0.002329660579562187, + 0.002516933251172304, + 0.005026699975132942, + 0.005296738818287849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0019507441902533174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644440021365881 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644440021365881, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644440021365881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2138500064611435, + 0.11621000617742538, + 0.10089410096406937, + 0.06352800130844116, + 0.07682575285434723, + 0.10961862653493881, + 0.15498505532741547, + 0.06064249202609062, + 0.07010192424058914, + 0.03334399312734604 + ], + "credit_map": [ + 0.0181503277271986, + 0.029536982998251915, + 0.017583854496479034, + 0.016347713768482208, + 0.020174801349639893, + 0.042466916143894196, + 0.019582895562052727, + 0.05000652000308037, + 0.00970045942813158, + 0.00995390024036169, + 0.019952991977334023, + 0.01579270511865616, + 0.010816128924489021, + 0.013857030309736729, + 0.014987786300480366, + 0.02114901877939701, + 0.009824685752391815, + 0.008441268466413021, + 0.020771505311131477, + 0.009900162927806377, + 0.010896429419517517, + 0.010519159026443958, + 0.012536357156932354, + 0.018004542216658592, + 0.00842229649424553, + 0.008719267323613167, + 0.007529987953603268, + 0.006407895591109991, + 0.008400284685194492, + 0.007843096740543842, + 0.006854583043605089, + 0.009350595064461231, + 0.008453143760561943, + 0.004936301615089178, + 0.01853913441300392, + 0.007249027956277132, + 0.00419456884264946, + 0.008276169188320637, + 0.007734397426247597, + 0.017443014308810234, + 0.009429983794689178, + 0.004455768968909979, + 0.0366009920835495, + 0.009061398915946484, + 0.004941694438457489, + 0.00791179295629263, + 0.010951614007353783, + 0.026265382766723633, + 0.008031994104385376, + 0.007766623981297016, + 0.03216350078582764, + 0.010078237392008305, + 0.006738980766385794, + 0.0071987127885222435, + 0.05416901409626007, + 0.028838010504841805, + 0.005704724695533514, + 0.007106452714651823, + 0.007999259978532791, + 0.0076085664331912994, + 0.006054742727428675, + 0.00480235181748867, + 0.01168547011911869, + 0.009680931456387043, + 0.007297905161976814, + 0.006796650122851133, + 0.01016717217862606, + 0.009689584374427795, + 0.006375869736075401, + 0.007365765515714884, + 0.011078575626015663, + 0.011330406181514263, + 0.004085347522050142, + 0.002973455237224698, + 0.004043919499963522, + 0.004247186239808798, + 0.0036610690876841545, + 0.002705981256440282, + 0.006361356470733881, + 0.005265680141746998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.004644440021365881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a vertical format with a close-up framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting steadily downward from a slight downward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brass pocket watch with intricate vintage patterns visible sliding down a deep burgundy silk handkerchief?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the brass watch case softly lit from the side, showing a warm metallic luster and slight wear on the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the watch slide down the handkerchief with a steady, deliberate motion that creates a shallow indentation in the fabric?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185662381350994 + ], + "opposite_logprob_tokens": [ + -4.264185428619385 + ], + "logprob_reward": -0.014185662381350994, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185662381350994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3437713086605072, + 0.11305148154497147, + 0.08294723927974701, + 0.12567950785160065, + 0.06769347935914993, + 0.06280852854251862, + 0.07707320153713226, + 0.039017435163259506, + 0.05784149095416069, + 0.030116289854049683 + ], + "credit_map": [ + 0.030415890738368034, + 0.056191232055425644, + 0.052452780306339264, + 0.030517857521772385, + 0.028707848861813545, + 0.056968796998262405, + 0.03504227101802826, + 0.053474653512239456, + 0.0092782536521554, + 0.012589425779879093, + 0.01969219744205475, + 0.020609524101018906, + 0.013121586292982101, + 0.012133605778217316, + 0.01187882386147976, + 0.013748073019087315, + 0.007519266102463007, + 0.008743534795939922, + 0.01306332927197218, + 0.009614909067749977, + 0.009054860100150108, + 0.010053454898297787, + 0.008695648983120918, + 0.01620223931968212, + 0.013759023509919643, + 0.01348737720400095, + 0.017159471288323402, + 0.011840743944048882, + 0.014700418338179588, + 0.020707255229353905, + 0.016786441206932068, + 0.017238784581422806, + 0.00866241380572319, + 0.004649953451007605, + 0.015082607977092266, + 0.00872413907200098, + 0.0038565220311284065, + 0.008173319511115551, + 0.00957565475255251, + 0.008968869224190712, + 0.005682819988578558, + 0.007857952266931534, + 0.01009515393525362, + 0.010538878850638866, + 0.004843680653721094, + 0.005115365143865347, + 0.010139909572899342, + 0.008534771390259266, + 0.007022721227258444, + 0.00475205248221755, + 0.023213235661387444, + 0.0059188841842114925, + 0.004168545827269554, + 0.003972024656832218, + 0.021196981891989708, + 0.006828757002949715, + 0.004247740842401981, + 0.003511920804157853, + 0.005540546029806137, + 0.0042370399460196495, + 0.002702867379412055, + 0.0021831884514540434, + 0.00866122916340828, + 0.007932904176414013, + 0.005109172314405441, + 0.0064029390923678875, + 0.007787749171257019, + 0.007282661274075508, + 0.0057382527738809586, + 0.004843615461140871, + 0.010866798460483551, + 0.009810301475226879, + 0.002716789720579982, + 0.0035929244477301836, + 0.0037797901313751936, + 0.003509339177981019, + 0.0030093027744442225, + 0.0033306204713881016, + 0.005385365802794695, + 0.004792158491909504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.014185662381350994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_tensor([7295, 7295])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_tensor([7295, 7295])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7f721a00e0efea9b524250fd27186c7cff53e1b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_tensor([7295, 7295])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "breathtaking slow-motion long shot, filmed from a low angle just above sea level, tilts upward to follow the breathtaking force of a massive humpback whale as it plunges from the deep blue ocean. Illuminated by the bright midday sun, its immense bulk traces a majestic arc through the air, its dark, slippery skin, sea-marked and speckled with white barnacles, gleaming in stark contrast to its light-colored, furrowed belly. Tons of seawater cascade from its body in a shimmering curtain of water and countless crystalline droplets, which catch the sunlight in a fleeting iridescent haze. At the height of its leap, its massive pectoral fins unfurl like wings, and its entire body appears to pause in the air for a moment before plummeting backward into the sea, unleashing a towering explosion of white spray and foam. The entire picture is full of the real texture of high-definition documentaries, capturing the majestic power and awe-inspiring beauty of this natural wonder, creating a grand and tranquil atmosphere.", + "seed": 4726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_7295_seed_4726_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0201061824336648, + "vlm_logprob": -0.0201061824336648, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0201061824336648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036501072347164154 + ], + "opposite_logprob_tokens": [ + -3.536501169204712 + ], + "logprob_reward": -0.036501072347164154, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.036501072347164154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.096431665122509, + 0.14760011434555054, + 0.23662734031677246, + 0.03242241218686104, + 0.27264532446861267, + 0.05127673223614693, + 0.08127366751432419, + 0.022064048796892166, + 0.020418552681803703, + 0.03924016281962395 + ], + "credit_map": [ + 0.01060810312628746, + 0.008437166921794415, + 0.008484626188874245, + 0.007200221996754408, + 0.01053757406771183, + 0.012788400053977966, + 0.01902606338262558, + 0.019349507987499237, + 0.01483978983014822, + 0.019291095435619354, + 0.018576445057988167, + 0.017154840752482414, + 0.019369876012206078, + 0.021184684708714485, + 0.02368900366127491, + 0.013494376093149185, + 0.01836252585053444, + 0.01069831196218729, + 0.015483387745916843, + 0.015017848461866379, + 0.08187152445316315, + 0.07413281500339508, + 0.009977553971111774, + 0.011083366349339485, + 0.005097950343042612, + 0.002836660947650671, + 0.003853581380099058, + 0.005672919098287821, + 0.003349147504195571, + 0.004584828857332468, + 0.0030319762881845236, + 0.003995346836745739, + 0.037023261189460754, + 0.018271055072546005, + 0.05300942435860634, + 0.02754150703549385, + 0.021281501278281212, + 0.054985981434583664, + 0.017790570855140686, + 0.04274202138185501, + 0.009045383892953396, + 0.005006204824894667, + 0.007955827750265598, + 0.006094053853303194, + 0.0034824316389858723, + 0.0079731997102499, + 0.0036574325058609247, + 0.008062195032835007, + 0.0076514314860105515, + 0.0076301489025354385, + 0.0178118534386158, + 0.005419437773525715, + 0.003734234720468521, + 0.015679484233260155, + 0.00668765464797616, + 0.01665942743420601, + 0.0020945763681083918, + 0.0019045061198994517, + 0.004580225329846144, + 0.001972703030332923, + 0.00243506976403296, + 0.0033956305123865604, + 0.003204658627510071, + 0.0024766777642071247, + 0.0030068038031458855, + 0.002540203742682934, + 0.003171829506754875, + 0.0020749210380017757, + 0.0016642905538901687, + 0.0028801155276596546, + 0.002191901206970215, + 0.002888487186282873, + 0.003065962577238679, + 0.0028144842945039272, + 0.012891964986920357, + 0.004393936600536108, + 0.0032861174549907446, + 0.007625034544616938, + 0.00254390318877995, + 0.0026187608018517494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.036501072347164154 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865588616579771 + ], + "opposite_logprob_tokens": [ + -4.758656024932861 + ], + "logprob_reward": -0.00865588616579771, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865588616579771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0902971401810646, + 0.04018072783946991, + 0.08567706495523453, + 0.12831918895244598, + 0.04073105379939079, + 0.0696658119559288, + 0.027746448293328285, + 0.037179671227931976, + 0.06524603068828583, + 0.4149568974971771 + ], + "credit_map": [ + 0.011569022201001644, + 0.01205044612288475, + 0.007947453297674656, + 0.00807948037981987, + 0.009096957743167877, + 0.007986437529325485, + 0.01865166798233986, + 0.014915675856173038, + 0.0041977115906775, + 0.003284791950136423, + 0.005033759865909815, + 0.004232867620885372, + 0.006181018427014351, + 0.006837720051407814, + 0.006475761532783508, + 0.003937094006687403, + 0.008063904009759426, + 0.005704165901988745, + 0.006157292053103447, + 0.009246996603906155, + 0.01627693697810173, + 0.018067779019474983, + 0.013887796550989151, + 0.008272193372249603, + 0.014754756353795528, + 0.01483702939003706, + 0.01591390185058117, + 0.020224379375576973, + 0.009793773293495178, + 0.01610085740685463, + 0.020107997581362724, + 0.01658649928867817, + 0.004836980253458023, + 0.0030642529018223286, + 0.006683106068521738, + 0.004471701569855213, + 0.003844753373414278, + 0.008012205362319946, + 0.003030465915799141, + 0.006787587888538837, + 0.011628802865743637, + 0.00823014136403799, + 0.010076789185404778, + 0.00604197196662426, + 0.007789211813360453, + 0.010042790323495865, + 0.004583330359309912, + 0.011272775009274483, + 0.003997109364718199, + 0.0033581096213310957, + 0.004113481845706701, + 0.00430975342169404, + 0.001185855595394969, + 0.003088289638981223, + 0.0014633182436227798, + 0.006230532191693783, + 0.003910129424184561, + 0.0028825581539422274, + 0.009597045369446278, + 0.003325999714434147, + 0.0031865728087723255, + 0.00506016006693244, + 0.004747211001813412, + 0.00446999492123723, + 0.004815819673240185, + 0.005247469991445541, + 0.02226092480123043, + 0.005088326521217823, + 0.003484075190499425, + 0.010436033830046654, + 0.003436693223193288, + 0.010476685129106045, + 0.02632463350892067, + 0.011773052625358105, + 0.17577910423278809, + 0.0380326509475708, + 0.011131424456834793, + 0.05185673385858536, + 0.005349904298782349, + 0.0947093814611435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.00865588616579771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595619454979897 + ], + "opposite_logprob_tokens": [ + -3.8955955505371094 + ], + "logprob_reward": -0.020595619454979897, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595619454979897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06950058043003082, + 0.10600556433200836, + 0.21562574803829193, + 0.08483288437128067, + 0.049989279359579086, + 0.0719190314412117, + 0.0712907686829567, + 0.050655297935009, + 0.1591491550207138, + 0.12103167921304703 + ], + "credit_map": [ + 0.008364420384168625, + 0.010387369431555271, + 0.012386772781610489, + 0.006041651125997305, + 0.00834787730127573, + 0.008323989808559418, + 0.005480288527905941, + 0.010168209671974182, + 0.01259545050561428, + 0.01679823361337185, + 0.017965978011488914, + 0.012518933974206448, + 0.010824699886143208, + 0.013353109359741211, + 0.01007043570280075, + 0.011878720484673977, + 0.0224147979170084, + 0.023087220266461372, + 0.025911875069141388, + 0.02338198572397232, + 0.03466431796550751, + 0.03374066576361656, + 0.03321409970521927, + 0.019210780039429665, + 0.009765944443643093, + 0.012035119347274303, + 0.008035936392843723, + 0.01323104277253151, + 0.008911671116948128, + 0.011100225150585175, + 0.011111507192254066, + 0.010641426779329777, + 0.004976640455424786, + 0.004474772606045008, + 0.007195373065769672, + 0.008146834559738636, + 0.005192875862121582, + 0.008138023316860199, + 0.004543352406471968, + 0.0073214066214859486, + 0.013545749709010124, + 0.006800371687859297, + 0.009768891148269176, + 0.008925515227019787, + 0.004436810035258532, + 0.011035009287297726, + 0.005317940842360258, + 0.01208873838186264, + 0.009525740519165993, + 0.007605258375406265, + 0.012608572840690613, + 0.006946103181689978, + 0.0030370817985385656, + 0.01161538902670145, + 0.005749350413680077, + 0.014203271828591824, + 0.006828356999903917, + 0.004416801501065493, + 0.00894211232662201, + 0.0073144095949828625, + 0.0039061992429196835, + 0.005948997102677822, + 0.006188470870256424, + 0.007109945639967918, + 0.01996282860636711, + 0.015535653568804264, + 0.023316558450460434, + 0.018627244979143143, + 0.014161357656121254, + 0.030484937131404877, + 0.013984630815684795, + 0.023075947538018227, + 0.009099657647311687, + 0.012111485935747623, + 0.0190036091953516, + 0.010854234918951988, + 0.011474129743874073, + 0.030362384393811226, + 0.007899864576756954, + 0.020226307213306427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.020595619454979897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199413359165192 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199413359165192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199413359165192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1563103199005127, + 0.05814789608120918, + 0.20440226793289185, + 0.11001317203044891, + 0.05624478682875633, + 0.045514777302742004, + 0.07573340088129044, + 0.07551004737615585, + 0.08313184231519699, + 0.13499152660369873 + ], + "credit_map": [ + 0.011784770525991917, + 0.015069644898176193, + 0.011407885700464249, + 0.013636223040521145, + 0.014417043887078762, + 0.015611034817993641, + 0.03624919056892395, + 0.03813451901078224, + 0.007408230099827051, + 0.009315090253949165, + 0.008680466562509537, + 0.007449198514223099, + 0.006049626041203737, + 0.007222254294902086, + 0.0058793434873223305, + 0.0061436863616108894, + 0.018525030463933945, + 0.016370121389627457, + 0.020328130573034286, + 0.02560986392199993, + 0.04167500510811806, + 0.04637991264462471, + 0.022119875997304916, + 0.013394329696893692, + 0.01214718259871006, + 0.012241884134709835, + 0.013397461734712124, + 0.01926608383655548, + 0.008919185027480125, + 0.013600315898656845, + 0.019374510273337364, + 0.011066540144383907, + 0.006010622717440128, + 0.004495515953749418, + 0.007185080088675022, + 0.007867860607802868, + 0.0059247626923024654, + 0.010135956108570099, + 0.00554549740627408, + 0.009079480543732643, + 0.008121195249259472, + 0.0054741548374295235, + 0.005761391017585993, + 0.005615344736725092, + 0.0038228975608944893, + 0.00613213749602437, + 0.0030503796879202127, + 0.007537274621427059, + 0.00836406834423542, + 0.008057577535510063, + 0.014713556505739689, + 0.007427187170833349, + 0.003743653418496251, + 0.011635713279247284, + 0.0045534162782132626, + 0.017238222062587738, + 0.016364743933081627, + 0.00421961210668087, + 0.017114490270614624, + 0.008662168867886066, + 0.004263630136847496, + 0.010498355142772198, + 0.005666360259056091, + 0.008720683865249157, + 0.009186905808746815, + 0.008686333894729614, + 0.018760493025183678, + 0.010136096738278866, + 0.007166608702391386, + 0.010995138436555862, + 0.007888746447861195, + 0.010311512276530266, + 0.009212406352162361, + 0.008897567167878151, + 0.04074908420443535, + 0.015699632465839386, + 0.006500151474028826, + 0.010142411105334759, + 0.004941280465573072, + 0.03884897008538246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.014199413359165192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057892084121704 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057892084121704, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057892084121704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21046270430088043, + 0.037210699170827866, + 0.09212139248847961, + 0.13356903195381165, + 0.09158185124397278, + 0.05053665488958359, + 0.06373389810323715, + 0.05684242025017738, + 0.12555252015590668, + 0.13838881254196167 + ], + "credit_map": [ + 0.016600914299488068, + 0.0195856224745512, + 0.015568360686302185, + 0.017000339925289154, + 0.021877169609069824, + 0.019899455830454826, + 0.0555327944457531, + 0.044398050755262375, + 0.00453069107607007, + 0.007514730095863342, + 0.0048122890293598175, + 0.004158513154834509, + 0.0037789347115904093, + 0.004585057031363249, + 0.004609831143170595, + 0.0032206519972532988, + 0.011049523949623108, + 0.0077955578453838825, + 0.012820151634514332, + 0.013570589944720268, + 0.010043435730040073, + 0.009678550995886326, + 0.0170535109937191, + 0.010110054165124893, + 0.015675866976380348, + 0.012312965467572212, + 0.015367269515991211, + 0.019660186022520065, + 0.01282304897904396, + 0.017433155328035355, + 0.019754210487008095, + 0.020542318001389503, + 0.010848615318536758, + 0.006699625868350267, + 0.015094734728336334, + 0.011226180009543896, + 0.009394554421305656, + 0.01615539938211441, + 0.007990729063749313, + 0.014172009192407131, + 0.009897278621792793, + 0.006243581417948008, + 0.006274825893342495, + 0.0072896541096270084, + 0.004377321805804968, + 0.006036075297743082, + 0.003766559297218919, + 0.006651357281953096, + 0.009057334624230862, + 0.004847317468374968, + 0.014794161543250084, + 0.007982113398611546, + 0.001988403731957078, + 0.009862011298537254, + 0.0035402150824666023, + 0.011662330478429794, + 0.0060343327932059765, + 0.005601696670055389, + 0.009699267335236073, + 0.006883018650114536, + 0.006315463688224554, + 0.006702780723571777, + 0.009149076417088509, + 0.006456778850406408, + 0.012542599812150002, + 0.009722638875246048, + 0.03324110805988312, + 0.012762101367115974, + 0.0062895407900214195, + 0.030457913875579834, + 0.009371072985231876, + 0.011165538802742958, + 0.008509734645485878, + 0.010775712318718433, + 0.039694059640169144, + 0.016010165214538574, + 0.006891402415931225, + 0.03839872032403946, + 0.004575286991894245, + 0.013533717021346092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02057892084121704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "breathtaking slow-motion long shot, filmed from a low angle just above sea level, tilts upward to follow the breathtaking force of a massive humpback whale as it plunges from the deep blue ocean. Illuminated by the bright midday sun, its immense bulk traces a majestic arc through the air, its dark, slippery skin, sea-marked and speckled with white barnacles, gleaming in stark contrast to its light-colored, furrowed belly. Tons of seawater cascade from its body in a shimmering curtain of water and countless crystalline droplets, which catch the sunlight in a fleeting iridescent haze. At the height of its leap, its massive pectoral fins unfurl like wings, and its entire body appears to pause in the air for a moment before plummeting backward into the sea, unleashing a towering explosion of white spray and foam. The entire picture is full of the real texture of high-definition documentaries, capturing the majestic power and awe-inspiring beauty of this natural wonder, creating a grand and tranquil atmosphere.", + "seed": 4727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_7295_seed_4727_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03654918074607849, + "vlm_logprob": -0.03654918074607849, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03654918074607849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02201313152909279 + ], + "opposite_logprob_tokens": [ + -4.147013187408447 + ], + "logprob_reward": -0.02201313152909279, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02201313152909279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2284785658121109, + 0.16683585941791534, + 0.025743715465068817, + 0.25084927678108215, + 0.05228783190250397, + 0.02077614516019821, + 0.03516833484172821, + 0.05293459817767143, + 0.10448611527681351, + 0.06243955343961716 + ], + "credit_map": [ + 0.032846853137016296, + 0.031199410557746887, + 0.03287127614021301, + 0.02602750062942505, + 0.022646890953183174, + 0.02942187339067459, + 0.022857438772916794, + 0.030607331544160843, + 0.015533216297626495, + 0.030215835198760033, + 0.019129443913698196, + 0.015508241020143032, + 0.029012339189648628, + 0.021962668746709824, + 0.01620263047516346, + 0.019271481782197952, + 0.002939100842922926, + 0.003386381082236767, + 0.0037563431542366743, + 0.002495815046131611, + 0.002486360492184758, + 0.00360304769128561, + 0.004460570868104696, + 0.0026160951238125563, + 0.035099439322948456, + 0.04334930703043938, + 0.03569042310118675, + 0.020838409662246704, + 0.03865710273385048, + 0.0313434936106205, + 0.022653255611658096, + 0.023217836394906044, + 0.005740801338106394, + 0.006452516652643681, + 0.0072843232192099094, + 0.0054862238466739655, + 0.007893659174442291, + 0.008822725154459476, + 0.0053161741234362125, + 0.005291410721838474, + 0.002223763382062316, + 0.0015391901833936572, + 0.00289731752127409, + 0.002705921418964863, + 0.002296390710398555, + 0.0018676645122468472, + 0.0024816482327878475, + 0.004764249082654715, + 0.003186544170603156, + 0.004284614231437445, + 0.007341019343584776, + 0.006809049751609564, + 0.004375583957880735, + 0.0038498167414218187, + 0.0030864542350172997, + 0.0022352521773427725, + 0.007568910717964172, + 0.009744090028107166, + 0.0067272004671394825, + 0.007087171543389559, + 0.005682953633368015, + 0.005820033606141806, + 0.00645658653229475, + 0.0038476502522826195, + 0.014871395193040371, + 0.012907183729112148, + 0.021908529102802277, + 0.016326379030942917, + 0.013140610419213772, + 0.007597052026540041, + 0.011357520706951618, + 0.006377449259161949, + 0.008047237060964108, + 0.008111611939966679, + 0.007739834953099489, + 0.007325374521315098, + 0.00827945675700903, + 0.01055729016661644, + 0.006354739889502525, + 0.006024009548127651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02201313152909279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709973096847534 + ], + "opposite_logprob_tokens": [ + -3.40871000289917 + ], + "logprob_reward": -0.033709973096847534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709973096847534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16203680634498596, + 0.15627597272396088, + 0.030033254995942116, + 0.058051615953445435, + 0.08287348598241806, + 0.11316724866628647, + 0.2211523801088333, + 0.05798230320215225, + 0.044532835483551025, + 0.07389413565397263 + ], + "credit_map": [ + 0.02089015021920204, + 0.023079300299286842, + 0.01667129248380661, + 0.024816129356622696, + 0.012753676623106003, + 0.019347207620739937, + 0.02137112244963646, + 0.023107919842004776, + 0.01821369118988514, + 0.020740985870361328, + 0.015013236552476883, + 0.012134527787566185, + 0.034506238996982574, + 0.014943672344088554, + 0.015115718357264996, + 0.02560790628194809, + 0.003044414333999157, + 0.003956858534365892, + 0.003063911572098732, + 0.004153999499976635, + 0.003446234157308936, + 0.0037351585924625397, + 0.004476545844227076, + 0.004156132694333792, + 0.008237795904278755, + 0.01285704504698515, + 0.007739976514130831, + 0.004999710712581873, + 0.005192231386899948, + 0.007564794272184372, + 0.005969875026494265, + 0.005490187089890242, + 0.00713867275044322, + 0.007973827421665192, + 0.015192201361060143, + 0.010929285548627377, + 0.011482231318950653, + 0.0122273163869977, + 0.00957306195050478, + 0.008356890641152859, + 0.01532624289393425, + 0.011528351344168186, + 0.011468038894236088, + 0.0160506684333086, + 0.01589149795472622, + 0.01450335793197155, + 0.012828854843974113, + 0.015570235438644886, + 0.025318261235952377, + 0.03277066722512245, + 0.03237326815724373, + 0.030004659667611122, + 0.03205813467502594, + 0.024923836812376976, + 0.017895430326461792, + 0.025808118283748627, + 0.008840510621666908, + 0.007291269954293966, + 0.009794222190976143, + 0.005164798349142075, + 0.008174806833267212, + 0.0053759426809847355, + 0.005306003615260124, + 0.008034750819206238, + 0.006629001349210739, + 0.004561253823339939, + 0.00801184494048357, + 0.00656856968998909, + 0.005578974727541208, + 0.004039566032588482, + 0.00470392033457756, + 0.004439703654497862, + 0.011408125050365925, + 0.007096217945218086, + 0.009267818182706833, + 0.008436757139861584, + 0.00816319976001978, + 0.010061659850180149, + 0.009306326508522034, + 0.010154036805033684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.033709973096847534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195417053997517 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195417053997517, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195417053997517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06963197141885757, + 0.3818930387496948, + 0.07690437138080597, + 0.08040637522935867, + 0.04325956106185913, + 0.0383676141500473, + 0.026134122163057327, + 0.15575723350048065, + 0.06146794185042381, + 0.06617776304483414 + ], + "credit_map": [ + 0.006922030355781317, + 0.009263062849640846, + 0.007936383597552776, + 0.00903795100748539, + 0.006829081568866968, + 0.008700700476765633, + 0.010438700206577778, + 0.010504075326025486, + 0.03030158206820488, + 0.10510087013244629, + 0.032468073070049286, + 0.032718054950237274, + 0.05634889006614685, + 0.036052118986845016, + 0.03440156579017639, + 0.05450190231204033, + 0.014328988268971443, + 0.009579659439623356, + 0.008452780544757843, + 0.010393223725259304, + 0.008421937935054302, + 0.007844921201467514, + 0.009459055960178375, + 0.008423811756074429, + 0.010875639505684376, + 0.0106605039909482, + 0.012396134436130524, + 0.006046118214726448, + 0.012819371186196804, + 0.011257029138505459, + 0.009475764818489552, + 0.006875821389257908, + 0.005386289209127426, + 0.005203062202781439, + 0.0062616681680083275, + 0.005122855771332979, + 0.0050774128176271915, + 0.006414995528757572, + 0.004677383694797754, + 0.005115893203765154, + 0.004400379955768585, + 0.005030347965657711, + 0.0038967765867710114, + 0.004975153133273125, + 0.006276324857026339, + 0.004813929088413715, + 0.005122038070112467, + 0.0038526649586856365, + 0.0021333005279302597, + 0.0021293701138347387, + 0.00699797784909606, + 0.00701066292822361, + 0.0018276333576068282, + 0.0014647998614236712, + 0.002705744933336973, + 0.0018646343378350139, + 0.019058283418416977, + 0.015291162766516209, + 0.01654849573969841, + 0.018830008804798126, + 0.021795902401208878, + 0.016027556732296944, + 0.01905483938753605, + 0.02915099821984768, + 0.009400862269103527, + 0.0056146481074392796, + 0.007334287744015455, + 0.011408943682909012, + 0.009492306038737297, + 0.005577013362199068, + 0.0060428897850215435, + 0.00659699784591794, + 0.007004887331277132, + 0.007974977605044842, + 0.007870460860431194, + 0.007980544120073318, + 0.009752395562827587, + 0.012976879253983498, + 0.006684597115963697, + 0.005933024454861879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.014195417053997517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029155761003494 + ], + "opposite_logprob_tokens": [ + -2.3502914905548096 + ], + "logprob_reward": -0.10029155761003494, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029155761003494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13633601367473602, + 0.0781039297580719, + 0.029874315485358238, + 0.04470541328191757, + 0.04682813957333565, + 0.057564627379179, + 0.24963653087615967, + 0.12024632841348648, + 0.18127870559692383, + 0.055425986647605896 + ], + "credit_map": [ + 0.015877339988946915, + 0.012626041658222675, + 0.018312592059373856, + 0.011849072761833668, + 0.020205501466989517, + 0.020860953256487846, + 0.013719757087528706, + 0.02288476563990116, + 0.007991096936166286, + 0.011307986453175545, + 0.01043783687055111, + 0.007651722989976406, + 0.01103308703750372, + 0.013737869448959827, + 0.007009586319327354, + 0.008934739977121353, + 0.005198642145842314, + 0.0034817857667803764, + 0.0040426054038107395, + 0.0030664971563965082, + 0.0028481159824877977, + 0.0035018562339246273, + 0.005129766184836626, + 0.0026050445158034563, + 0.004600649233907461, + 0.006543806288391352, + 0.008069338276982307, + 0.0037974228616803885, + 0.005466016940772533, + 0.007064327131956816, + 0.0060994625091552734, + 0.0030643886420875788, + 0.0057615600526332855, + 0.00497540133073926, + 0.00897972472012043, + 0.005781054496765137, + 0.004828659817576408, + 0.00824868306517601, + 0.004237265791743994, + 0.004015791229903698, + 0.007325344253331423, + 0.006198677700012922, + 0.00694234948605299, + 0.007673061918467283, + 0.0084880655631423, + 0.00718092592433095, + 0.006132282316684723, + 0.007623919751495123, + 0.025349203497171402, + 0.02792276069521904, + 0.05340452492237091, + 0.057429295033216476, + 0.02286747470498085, + 0.02024681679904461, + 0.02262214571237564, + 0.019794316962361336, + 0.01694553904235363, + 0.02364516817033291, + 0.018988728523254395, + 0.013882225379347801, + 0.010081551037728786, + 0.016092175617814064, + 0.007436713669449091, + 0.013174226507544518, + 0.03286406770348549, + 0.017304187640547752, + 0.02886335179209709, + 0.022854747250676155, + 0.01228137593716383, + 0.015050061047077179, + 0.025450395420193672, + 0.026610521599650383, + 0.008072622120380402, + 0.006348993629217148, + 0.007874983362853527, + 0.0069140722043812275, + 0.006696744356304407, + 0.006148362532258034, + 0.0075412834994494915, + 0.0058289263397455215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.10029155761003494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26535218954086304, + 0.1783704310655594, + 0.023624662309885025, + 0.036179445683956146, + 0.058578554540872574, + 0.06928860396146774, + 0.08799251914024353, + 0.1464654952287674, + 0.06583162397146225, + 0.06831658631563187 + ], + "credit_map": [ + 0.03701109066605568, + 0.025625817477703094, + 0.03518494963645935, + 0.024825938045978546, + 0.03878765553236008, + 0.03996945172548294, + 0.020311471074819565, + 0.04363580793142319, + 0.013435824774205685, + 0.04094638302922249, + 0.016371896490454674, + 0.01326686143875122, + 0.0317460261285305, + 0.017165835946798325, + 0.01814717799425125, + 0.02729041874408722, + 0.002605853835120797, + 0.0032526999711990356, + 0.0025566969998180866, + 0.0028706216253340244, + 0.003015137044712901, + 0.002451606560498476, + 0.004546638578176498, + 0.0023254083935171366, + 0.004687213338911533, + 0.007804526947438717, + 0.004229155369102955, + 0.0032308220397681, + 0.004995314870029688, + 0.0039237989112734795, + 0.003667088458314538, + 0.0036415262147784233, + 0.005067941267043352, + 0.007455192971974611, + 0.00686231954023242, + 0.007478912826627493, + 0.010026942007243633, + 0.006175008602440357, + 0.007605448830872774, + 0.007906782440841198, + 0.009054426103830338, + 0.007297849282622337, + 0.0072816479951143265, + 0.009831039234995842, + 0.00940324179828167, + 0.008564506657421589, + 0.007781547959893942, + 0.010074338875710964, + 0.008972708135843277, + 0.014221982099115849, + 0.00974876619875431, + 0.011991997249424458, + 0.013711546547710896, + 0.010446619242429733, + 0.007856872864067554, + 0.011042024940252304, + 0.018278278410434723, + 0.011215831153094769, + 0.019483240321278572, + 0.02195928990840912, + 0.01626453921198845, + 0.012259166687726974, + 0.017441602423787117, + 0.029563549906015396, + 0.007782563101500273, + 0.00493637565523386, + 0.007686124183237553, + 0.011925940401852131, + 0.010718070901930332, + 0.005174305289983749, + 0.0086311474442482, + 0.008977090939879417, + 0.008558517321944237, + 0.008567585609853268, + 0.00876349862664938, + 0.008189927786588669, + 0.007780231535434723, + 0.01164268795400858, + 0.008002245798707008, + 0.006811895407736301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012535824440419674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "breathtaking slow-motion long shot, filmed from a low angle just above sea level, tilts upward to follow the breathtaking force of a massive humpback whale as it plunges from the deep blue ocean. Illuminated by the bright midday sun, its immense bulk traces a majestic arc through the air, its dark, slippery skin, sea-marked and speckled with white barnacles, gleaming in stark contrast to its light-colored, furrowed belly. Tons of seawater cascade from its body in a shimmering curtain of water and countless crystalline droplets, which catch the sunlight in a fleeting iridescent haze. At the height of its leap, its massive pectoral fins unfurl like wings, and its entire body appears to pause in the air for a moment before plummeting backward into the sea, unleashing a towering explosion of white spray and foam. The entire picture is full of the real texture of high-definition documentaries, capturing the majestic power and awe-inspiring beauty of this natural wonder, creating a grand and tranquil atmosphere.", + "seed": 4728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_7295_seed_4728_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011987154744565487, + "vlm_logprob": -0.011987154744565487, + "vlm_yes_no_margin": 4.624999952316284, + "vlm_token_credit": -0.011987154744565487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020153163000941277 + ], + "opposite_logprob_tokens": [ + -4.145153045654297 + ], + "logprob_reward": -0.020153163000941277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020153163000941277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21229486167430878, + 0.3479295074939728, + 0.179719939827919, + 0.027430439367890358, + 0.015189085155725479, + 0.042635414749383926, + 0.07056735455989838, + 0.018180856481194496, + 0.051286403089761734, + 0.034766122698783875 + ], + "credit_map": [ + 0.030475547537207603, + 0.02493143081665039, + 0.025004539638757706, + 0.027029577642679214, + 0.0377633161842823, + 0.03031097538769245, + 0.022959614172577858, + 0.013819865882396698, + 0.015964286401867867, + 0.014221462421119213, + 0.03820928558707237, + 0.020365945994853973, + 0.023381181061267853, + 0.12480146437883377, + 0.053954508155584335, + 0.05703140050172806, + 0.011656609363853931, + 0.026304056867957115, + 0.02154933102428913, + 0.015140552073717117, + 0.031696073710918427, + 0.024066545069217682, + 0.028849322348833084, + 0.020457440987229347, + 0.003351567080244422, + 0.002695143921300769, + 0.0020832354202866554, + 0.0029270160011947155, + 0.0028592715971171856, + 0.004141585901379585, + 0.005806120578199625, + 0.003566497005522251, + 0.002078166464343667, + 0.0023021292872726917, + 0.0016256863018497825, + 0.0013285840395838022, + 0.00283768679946661, + 0.0021906422916799784, + 0.0015622804639860988, + 0.001263909274712205, + 0.006679986137896776, + 0.00791076198220253, + 0.005861273966729641, + 0.0031344217713922262, + 0.00550088519230485, + 0.00564995501190424, + 0.003666596719995141, + 0.00423153443261981, + 0.008404193446040154, + 0.008653557859361172, + 0.008154473267495632, + 0.0074823712930083275, + 0.00852744746953249, + 0.01127879973500967, + 0.009405748918652534, + 0.008660759776830673, + 0.0021294942125678062, + 0.0024812929332256317, + 0.0025592087768018246, + 0.0019004713976755738, + 0.002563662827014923, + 0.0026827824767678976, + 0.002039518905803561, + 0.0018244247185066342, + 0.006419460289180279, + 0.005850409157574177, + 0.0148426853120327, + 0.006880557630211115, + 0.004808926489204168, + 0.0027350110467523336, + 0.006652049720287323, + 0.0030973039101809263, + 0.003433411242440343, + 0.006979530211538076, + 0.0023552668280899525, + 0.0016177879879251122, + 0.003244611667469144, + 0.01230746228247881, + 0.0022021078038960695, + 0.0026259454898536205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.020153163000941277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108570583164692 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.01108570583164692, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108570583164692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052289996296167374, + 0.19511756300926208, + 0.05950778350234032, + 0.04422971233725548, + 0.015796560794115067, + 0.06746082752943039, + 0.14647747576236725, + 0.06191551685333252, + 0.10754279792308807, + 0.24966178834438324 + ], + "credit_map": [ + 0.004141028970479965, + 0.004680011421442032, + 0.0074603664688766, + 0.006582244765013456, + 0.009985889308154583, + 0.008265160955488682, + 0.006390183232724667, + 0.0047851125709712505, + 0.009245997294783592, + 0.007331137079745531, + 0.009622632525861263, + 0.008252479135990143, + 0.011210923083126545, + 0.06927058100700378, + 0.027323219925165176, + 0.05286058411002159, + 0.0053573171608150005, + 0.007702077738940716, + 0.00676836259663105, + 0.006290616001933813, + 0.006849640980362892, + 0.009307634085416794, + 0.005784300621598959, + 0.011447837576270103, + 0.0039857784286141396, + 0.003191067138686776, + 0.008506973274052143, + 0.0031017225701361895, + 0.003133389400318265, + 0.010162806138396263, + 0.004804594907909632, + 0.007343381177634001, + 0.0019129026914015412, + 0.0017323096981272101, + 0.0013351801317185163, + 0.0014899778179824352, + 0.0028221197426319122, + 0.00215164665132761, + 0.001983245834708214, + 0.0023691782262176275, + 0.011036421172320843, + 0.012310695834457874, + 0.008644459769129753, + 0.004873157013207674, + 0.010771694593131542, + 0.007349529769271612, + 0.005543775390833616, + 0.006931083742529154, + 0.02033296599984169, + 0.018380455672740936, + 0.023748159408569336, + 0.023725736886262894, + 0.01535048708319664, + 0.016986824572086334, + 0.013061203062534332, + 0.014891642145812511, + 0.007057803217321634, + 0.009006445296108723, + 0.012157772667706013, + 0.00920054130256176, + 0.004384413827210665, + 0.007368538063019514, + 0.0069766901433467865, + 0.005763309542089701, + 0.016385015100240707, + 0.016768086701631546, + 0.02755332924425602, + 0.013269818387925625, + 0.01054218877106905, + 0.005889067426323891, + 0.011101474985480309, + 0.006033815443515778, + 0.02513500489294529, + 0.056584879755973816, + 0.016989178955554962, + 0.006737956777215004, + 0.02149089053273201, + 0.1066519245505333, + 0.005703470669686794, + 0.010368481278419495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.01108570583164692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05822223797440529, + 0.2293984591960907, + 0.16065619885921478, + 0.05432552471756935, + 0.06313157081604004, + 0.04592357203364372, + 0.09325350075960159, + 0.047387685626745224, + 0.04410702735185623, + 0.20359426736831665 + ], + "credit_map": [ + 0.005040621850639582, + 0.006578501313924789, + 0.009497914463281631, + 0.0075529771856963634, + 0.010139279067516327, + 0.007583817467093468, + 0.007273297756910324, + 0.004555822350084782, + 0.012011134065687656, + 0.01670987159013748, + 0.03590640425682068, + 0.02085789106786251, + 0.020412717014551163, + 0.04845296964049339, + 0.04552142322063446, + 0.029526010155677795, + 0.01211188267916441, + 0.021700752899050713, + 0.01948537677526474, + 0.014201962389051914, + 0.023392021656036377, + 0.01880054362118244, + 0.023946279659867287, + 0.027017351239919662, + 0.006361484527587891, + 0.006000758148729801, + 0.004117610864341259, + 0.00637388788163662, + 0.005428161472082138, + 0.006868692580610514, + 0.011746846139431, + 0.00742807611823082, + 0.008228770457208157, + 0.010084863752126694, + 0.005434670951217413, + 0.005909593775868416, + 0.012646574527025223, + 0.006822633557021618, + 0.006258287001401186, + 0.007746173068881035, + 0.005580518394708633, + 0.006987025961279869, + 0.005521605256944895, + 0.004191463813185692, + 0.008695882745087147, + 0.003604431403800845, + 0.006393984891474247, + 0.0049486542120575905, + 0.013028561137616634, + 0.011486988514661789, + 0.014896883629262447, + 0.018140239641070366, + 0.008533281274139881, + 0.009489044547080994, + 0.009277932345867157, + 0.008400559425354004, + 0.00556547986343503, + 0.007106448989361525, + 0.009060666896402836, + 0.006993400864303112, + 0.004257929977029562, + 0.005185570102185011, + 0.005886558908969164, + 0.0033316269982606173, + 0.007160861510783434, + 0.007089458871632814, + 0.004983667749911547, + 0.004576866514980793, + 0.004918103571981192, + 0.0036670982372015715, + 0.008337168022990227, + 0.0033737963531166315, + 0.019560590386390686, + 0.044694285839796066, + 0.016040975227952003, + 0.006710662506520748, + 0.0169675275683403, + 0.08612548559904099, + 0.005347844213247299, + 0.008146882057189941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04104718193411827, + 0.43882104754447937, + 0.018566295504570007, + 0.029161706566810608, + 0.008638077415525913, + 0.04833555966615677, + 0.11073808372020721, + 0.0201830193400383, + 0.06959119439125061, + 0.214917853474617 + ], + "credit_map": [ + 0.00331512070260942, + 0.004841455724090338, + 0.004536367487162352, + 0.004291406832635403, + 0.00793339591473341, + 0.007100678514689207, + 0.004022907931357622, + 0.005005848594009876, + 0.021133918315172195, + 0.01651211641728878, + 0.04891624674201012, + 0.02413889206945896, + 0.027767715975642204, + 0.15641149878501892, + 0.08229480683803558, + 0.06164585426449776, + 0.0015245568938553333, + 0.0019659141544252634, + 0.00208447128534317, + 0.0018674462335184216, + 0.0022715763188898563, + 0.002467015292495489, + 0.0019742376171052456, + 0.004411076195538044, + 0.0038381381891667843, + 0.0023867860436439514, + 0.004530155565589666, + 0.0025862567126750946, + 0.002644346095621586, + 0.006250997073948383, + 0.004336570389568806, + 0.0025884564965963364, + 0.0010403355117887259, + 0.0011551666539162397, + 0.0010829160455614328, + 0.000923479616176337, + 0.0015788525342941284, + 0.0011014350457116961, + 0.000834536156617105, + 0.0009213557932525873, + 0.0081599997356534, + 0.009109258651733398, + 0.006353187374770641, + 0.0035398914478719234, + 0.006945999804884195, + 0.00475862342864275, + 0.004324797540903091, + 0.005143801216036081, + 0.01681632362306118, + 0.014335471205413342, + 0.020540039986371994, + 0.019793560728430748, + 0.010456251911818981, + 0.00974365882575512, + 0.009370258077979088, + 0.009682519361376762, + 0.0020910354796797037, + 0.0034313835203647614, + 0.005212823860347271, + 0.0023154765367507935, + 0.0018502987222746015, + 0.0020248752553015947, + 0.0018595923902466893, + 0.0013975335750728846, + 0.00734295416623354, + 0.008683926425874233, + 0.017591921612620354, + 0.00793395284563303, + 0.007327516097575426, + 0.0045890505425632, + 0.009719066321849823, + 0.00640280032530427, + 0.02168995887041092, + 0.04859330132603645, + 0.014331837184727192, + 0.006030161399394274, + 0.01834186166524887, + 0.09179265797138214, + 0.005018024705350399, + 0.009120038710534573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.01818765513598919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03628982976078987, + 0.06553677469491959, + 0.03546784073114395, + 0.010063999332487583, + 0.013099980540573597, + 0.030096299946308136, + 0.047280654311180115, + 0.02054627053439617, + 0.11083805561065674, + 0.6307803392410278 + ], + "credit_map": [ + 0.0038351058028638363, + 0.003641183255240321, + 0.005168745759874582, + 0.005364228971302509, + 0.005779826547950506, + 0.005086973775178194, + 0.00416019419208169, + 0.0032535733189433813, + 0.004174193367362022, + 0.006032617297023535, + 0.00630256487056613, + 0.007986700162291527, + 0.007119892165064812, + 0.016780944541096687, + 0.005104632116854191, + 0.012035232968628407, + 0.00247257761657238, + 0.006054060533642769, + 0.003729932475835085, + 0.002864358713850379, + 0.006648761220276356, + 0.0038097272627055645, + 0.005472863558679819, + 0.004415561445057392, + 0.0009598543401807547, + 0.0009471143712289631, + 0.0021036018151789904, + 0.0011895184870809317, + 0.0006437677075155079, + 0.001865197322331369, + 0.0008390297298319638, + 0.0015159155009314418, + 0.0013385178754106164, + 0.0014612157829105854, + 0.0011736575979739428, + 0.001180312130600214, + 0.0029788922984153032, + 0.001643408089876175, + 0.0015629236586391926, + 0.0017610529903322458, + 0.004624273162335157, + 0.0052944994531571865, + 0.004111107904464006, + 0.0022800927981734276, + 0.004527389071881771, + 0.003276197938248515, + 0.0028727101162075996, + 0.0031100292690098286, + 0.00792973767966032, + 0.006774150300770998, + 0.0072494312189519405, + 0.00750009436160326, + 0.0034717805683612823, + 0.0038799240719527006, + 0.004767648875713348, + 0.005707884673029184, + 0.0023060310631990433, + 0.0025126198306679726, + 0.003424222581088543, + 0.0022981527727097273, + 0.0026290847454220057, + 0.0028572618030011654, + 0.002468830905854702, + 0.0020500675309449434, + 0.014790215529501438, + 0.012316807173192501, + 0.013646334409713745, + 0.014222162775695324, + 0.012206577695906162, + 0.01190237421542406, + 0.02090766467154026, + 0.010845919139683247, + 0.0641103982925415, + 0.1408700942993164, + 0.04366394132375717, + 0.017337512224912643, + 0.054068077355623245, + 0.2689725458621979, + 0.015751922503113747, + 0.026005838066339493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.005250830203294754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "breathtaking slow-motion long shot, filmed from a low angle just above sea level, tilts upward to follow the breathtaking force of a massive humpback whale as it plunges from the deep blue ocean. Illuminated by the bright midday sun, its immense bulk traces a majestic arc through the air, its dark, slippery skin, sea-marked and speckled with white barnacles, gleaming in stark contrast to its light-colored, furrowed belly. Tons of seawater cascade from its body in a shimmering curtain of water and countless crystalline droplets, which catch the sunlight in a fleeting iridescent haze. At the height of its leap, its massive pectoral fins unfurl like wings, and its entire body appears to pause in the air for a moment before plummeting backward into the sea, unleashing a towering explosion of white spray and foam. The entire picture is full of the real texture of high-definition documentaries, capturing the majestic power and awe-inspiring beauty of this natural wonder, creating a grand and tranquil atmosphere.", + "seed": 4729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000036/group_9/index_7295_seed_4729_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02893318235874176, + "vlm_logprob": -0.02893318235874176, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.02893318235874176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0538664273917675 + ], + "opposite_logprob_tokens": [ + -3.053866386413574 + ], + "logprob_reward": -0.0538664273917675, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0538664273917675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20157073438167572, + 0.13659335672855377, + 0.1859673261642456, + 0.059674035757780075, + 0.10884706676006317, + 0.08218851685523987, + 0.029509784653782845, + 0.050770387053489685, + 0.04813887178897858, + 0.09673983603715897 + ], + "credit_map": [ + 0.012153060175478458, + 0.07189605385065079, + 0.007968143559992313, + 0.009425939060747623, + 0.020999319851398468, + 0.0599176250398159, + 0.01010048110038042, + 0.00911013688892126, + 0.016229605302214622, + 0.021869536489248276, + 0.01232676487416029, + 0.016780003905296326, + 0.01994413696229458, + 0.01604372262954712, + 0.016807863488793373, + 0.016591740772128105, + 0.01647769846022129, + 0.02278292365372181, + 0.026581062003970146, + 0.032977182418107986, + 0.018631301820278168, + 0.020344585180282593, + 0.027854522690176964, + 0.020318059250712395, + 0.008108382113277912, + 0.009660483337938786, + 0.007684455718845129, + 0.005517813842743635, + 0.008809817023575306, + 0.0076293726451694965, + 0.005974228959530592, + 0.006289487238973379, + 0.012576736509799957, + 0.010781521908938885, + 0.018690306693315506, + 0.02328328602015972, + 0.011596628464758396, + 0.013088504783809185, + 0.01023848820477724, + 0.008591600693762302, + 0.00971518736332655, + 0.00876686256378889, + 0.012896653264760971, + 0.01403567474335432, + 0.008043240755796432, + 0.011377120390534401, + 0.008140435442328453, + 0.009213349781930447, + 0.004853629507124424, + 0.004492352716624737, + 0.004620512947440147, + 0.004680973012000322, + 0.0036402393598109484, + 0.0029696463607251644, + 0.0020379386842250824, + 0.0022144943941384554, + 0.006288126111030579, + 0.005292203277349472, + 0.005314817652106285, + 0.0066236406564712524, + 0.007719122339040041, + 0.00820751953870058, + 0.005093851592391729, + 0.006231110077351332, + 0.005632035434246063, + 0.004948321729898453, + 0.0048707881942391396, + 0.010180016048252583, + 0.006104358471930027, + 0.00743426289409399, + 0.005181512329727411, + 0.0037875801790505648, + 0.010888397693634033, + 0.012237554416060448, + 0.011850854381918907, + 0.013142701238393784, + 0.009703264571726322, + 0.009622753597795963, + 0.008681542240083218, + 0.020612778142094612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0538664273917675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033718038350343704 + ], + "opposite_logprob_tokens": [ + -3.4087181091308594 + ], + "logprob_reward": -0.033718038350343704, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033718038350343704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13367170095443726, + 0.3009468615055084, + 0.13599920272827148, + 0.07274089753627777, + 0.06253078579902649, + 0.039191991090774536, + 0.09473998099565506, + 0.052409712225198746, + 0.03967948630452156, + 0.06808938086032867 + ], + "credit_map": [ + 0.009496551938354969, + 0.03742159157991409, + 0.006683477200567722, + 0.006704626139253378, + 0.011026296764612198, + 0.047249581664800644, + 0.009457304142415524, + 0.0056322733871638775, + 0.0349666029214859, + 0.041683379560709, + 0.02772851102054119, + 0.04188118502497673, + 0.046029817312955856, + 0.03471881151199341, + 0.03627249225974083, + 0.03766607120633125, + 0.014540101401507854, + 0.01606246829032898, + 0.014873485080897808, + 0.021113337948918343, + 0.014528821222484112, + 0.012672684155404568, + 0.02791125886142254, + 0.014297053217887878, + 0.01001626718789339, + 0.010328900068998337, + 0.00831455085426569, + 0.0076726507395505905, + 0.011046318337321281, + 0.01108595635741949, + 0.00573158822953701, + 0.008544666692614555, + 0.007724490948021412, + 0.008050081320106983, + 0.009004423394799232, + 0.01123872958123684, + 0.005075592547655106, + 0.007815591990947723, + 0.006892671808600426, + 0.006729204207658768, + 0.005251773167401552, + 0.006309386342763901, + 0.006514032371342182, + 0.004597344435751438, + 0.004200695548206568, + 0.004697494674474001, + 0.0045625935308635235, + 0.003058672184124589, + 0.01362917385995388, + 0.010261885821819305, + 0.017616031691432, + 0.01636599749326706, + 0.009399281814694405, + 0.010774784721434116, + 0.00815594382584095, + 0.0085368687286973, + 0.009397994726896286, + 0.005454170051962137, + 0.009026889689266682, + 0.007219623774290085, + 0.005954431835561991, + 0.005492225289344788, + 0.004538498353213072, + 0.005325880832970142, + 0.006626902613788843, + 0.003928521182388067, + 0.005023299250751734, + 0.005859987810254097, + 0.0042956313118338585, + 0.004612015560269356, + 0.004101434722542763, + 0.0052316938526928425, + 0.008000020869076252, + 0.005913425236940384, + 0.009577116928994656, + 0.012345701456069946, + 0.006346424575895071, + 0.00751413032412529, + 0.005750962533056736, + 0.012641597539186478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.033718038350343704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549597769975662 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549597769975662, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549597769975662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4783795177936554, + 0.10931653529405594, + 0.028006289154291153, + 0.038511972874403, + 0.014773712493479252, + 0.06288781762123108, + 0.03745134919881821, + 0.0589669868350029, + 0.03429181128740311, + 0.13741402328014374 + ], + "credit_map": [ + 0.02448519505560398, + 0.18851874768733978, + 0.014322134666144848, + 0.01574729010462761, + 0.045928385108709335, + 0.16039660573005676, + 0.015295869670808315, + 0.01368526741862297, + 0.012831520289182663, + 0.012509332969784737, + 0.01021825335919857, + 0.016356807202100754, + 0.015431812033057213, + 0.011626249179244041, + 0.015963532030582428, + 0.014379027299582958, + 0.0032367766834795475, + 0.0034442851319909096, + 0.0034956499002873898, + 0.0045508574694395065, + 0.0034224707633256912, + 0.002746732672676444, + 0.004484580829739571, + 0.002624934772029519, + 0.005365792661905289, + 0.0057171485386788845, + 0.004293516743928194, + 0.003598716575652361, + 0.00710546039044857, + 0.005461609922349453, + 0.003708783071488142, + 0.003260943340137601, + 0.0019371169619262218, + 0.001969829900190234, + 0.002596848411485553, + 0.001998305320739746, + 0.0010991625022143126, + 0.0019675104413181543, + 0.0014516019728034735, + 0.0017533365171402693, + 0.007801962085068226, + 0.007114930544048548, + 0.010819214396178722, + 0.010064923204481602, + 0.005587676540017128, + 0.008789640851318836, + 0.0061705750413239, + 0.0065388926304876804, + 0.005286224652081728, + 0.005282546393573284, + 0.005158195272088051, + 0.0073469760827720165, + 0.0046360245905816555, + 0.003984365612268448, + 0.002762823598459363, + 0.0029941911343485117, + 0.01279132254421711, + 0.00438811257481575, + 0.010783662088215351, + 0.009303173050284386, + 0.006104352418333292, + 0.005524988751858473, + 0.004189575556665659, + 0.005881795194000006, + 0.006011671386659145, + 0.0035713829565793276, + 0.004424004349857569, + 0.006281111389398575, + 0.0036159525625407696, + 0.003949415870010853, + 0.002270085271447897, + 0.0041681877337396145, + 0.01501691434532404, + 0.016337523236870766, + 0.018205178901553154, + 0.018196137621998787, + 0.011792930774390697, + 0.013840561732649803, + 0.011801555752754211, + 0.03222322091460228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012549597769975662 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347821578383446 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347821578383446, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347821578383446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4303273856639862, + 0.07902785390615463, + 0.06161313131451607, + 0.040079087018966675, + 0.09215995669364929, + 0.029974315315485, + 0.13537852466106415, + 0.05125189200043678, + 0.04413178190588951, + 0.03605613112449646 + ], + "credit_map": [ + 0.023201314732432365, + 0.15800859034061432, + 0.016669679433107376, + 0.019469192251563072, + 0.04260993376374245, + 0.13401703536510468, + 0.01862769015133381, + 0.017723923549056053, + 0.006989330518990755, + 0.0098250238224864, + 0.007544189225882292, + 0.01706479676067829, + 0.01107484195381403, + 0.010243670083582401, + 0.009989834390580654, + 0.00629615830257535, + 0.007025793194770813, + 0.007953631691634655, + 0.008797228336334229, + 0.009408121928572655, + 0.008786842226982117, + 0.007007598411291838, + 0.005907565820962191, + 0.006726343184709549, + 0.004972726572304964, + 0.004839520901441574, + 0.005427825730293989, + 0.0038516675122082233, + 0.005872166715562344, + 0.006354945711791515, + 0.004733402747660875, + 0.004026828799396753, + 0.01158975251019001, + 0.01022170577198267, + 0.015341666527092457, + 0.018823523074388504, + 0.00891038216650486, + 0.010888667777180672, + 0.00831535179167986, + 0.008068897761404514, + 0.003698542481288314, + 0.004847372882068157, + 0.004931400530040264, + 0.00545995868742466, + 0.0025906264781951904, + 0.0029318344313651323, + 0.002719176933169365, + 0.0027953998651355505, + 0.018896710127592087, + 0.01703580468893051, + 0.021136479452252388, + 0.025719748809933662, + 0.019261015579104424, + 0.013055993244051933, + 0.010863911360502243, + 0.009408853948116302, + 0.010613434948027134, + 0.004916922189295292, + 0.007684886921197176, + 0.008500214666128159, + 0.006480613723397255, + 0.005386253818869591, + 0.0033512189984321594, + 0.004318338353186846, + 0.008684077300131321, + 0.005572826601564884, + 0.0064792511984705925, + 0.006533768493682146, + 0.004342674743384123, + 0.0040471190586686134, + 0.00333937699906528, + 0.005132684018462896, + 0.003621308133006096, + 0.002721573691815138, + 0.0049311197362840176, + 0.0055376640520989895, + 0.0020433079916983843, + 0.0054452912881970406, + 0.002448089886456728, + 0.009307772852480412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.026347821578383446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale shown plunging from the deep blue ocean into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the whale's skin appear dark and slippery with white barnacles and a light-colored furrowed belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera filmed from a low angle just above sea level and tilting upward to follow the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show tons of seawater cascading from the whale in a shimmering curtain of droplets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184026703238487 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184026703238487, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184026703238487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2719927728176117, + 0.26512521505355835, + 0.07838854938745499, + 0.05253171920776367, + 0.029716769233345985, + 0.032652802765369415, + 0.14046426117420197, + 0.055898234248161316, + 0.033267050981521606, + 0.03996261954307556 + ], + "credit_map": [ + 0.019743071869015694, + 0.09156841039657593, + 0.01391295250505209, + 0.015676509588956833, + 0.03825996443629265, + 0.06102948263287544, + 0.017134718596935272, + 0.014667670242488384, + 0.031084543094038963, + 0.03125504031777382, + 0.026519635692238808, + 0.040092747658491135, + 0.03752152621746063, + 0.027714790776371956, + 0.03610991686582565, + 0.03482703119516373, + 0.00680655287578702, + 0.01077183522284031, + 0.00771030317991972, + 0.009268865920603275, + 0.007971555925905704, + 0.00910329818725586, + 0.017172954976558685, + 0.009583184495568275, + 0.00908081978559494, + 0.007818756625056267, + 0.007172415032982826, + 0.003566762199625373, + 0.007363142911344767, + 0.007601289544254541, + 0.004355486016720533, + 0.005573046859353781, + 0.0036408083979040384, + 0.003433857811614871, + 0.0044901445508003235, + 0.005622098688036203, + 0.002497437410056591, + 0.003690628567710519, + 0.003533286741003394, + 0.0028085079975426197, + 0.004204131662845612, + 0.004224847536534071, + 0.004739319905638695, + 0.006101895123720169, + 0.003453803714364767, + 0.0038294096011668444, + 0.0030615730211138725, + 0.0030378217343240976, + 0.019402235746383667, + 0.01547219417989254, + 0.02375098690390587, + 0.02503710798919201, + 0.015740662813186646, + 0.016684938222169876, + 0.01255045086145401, + 0.011825690977275372, + 0.012039482593536377, + 0.004170789383351803, + 0.010484127327799797, + 0.008264384232461452, + 0.006378557998687029, + 0.00580248748883605, + 0.0026616104878485203, + 0.006096797063946724, + 0.006894283927977085, + 0.003779937280341983, + 0.005072187166661024, + 0.005796995013952255, + 0.002361117862164974, + 0.003102205228060484, + 0.002616910496726632, + 0.003643415402621031, + 0.004791239742189646, + 0.005127286538481712, + 0.004723657853901386, + 0.007297224830836058, + 0.003434739774093032, + 0.004816957283765078, + 0.003648338606581092, + 0.006123173516243696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.018184026703238487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_tensor([278, 278])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_tensor([278, 278])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90806c45081698df6fc6d2020db3c8bb344da032 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_tensor([278, 278])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, filmed from a low angle, the camera tilts upwards at a slow, steady pace, tracking a sudden commotion. A vast expanse of wilderness, the ground covered with withered, dry grass, dotted with a few twisted, leafless trees. The sky is shrouded in thick, swirling leaden clouds, and the faint, diffuse light casts an oppressive, monochromatic hue. Hundreds, perhaps thousands, of pitch-black crows, their feathers gleaming a glossy blue-purple in the gloom, roost on the ground and on dead branches. As if startled by an invisible, resounding roar, they suddenly flap their wings, emitting a chaotic, piercing cawing noise, and like an exploding black wave, they surge into the gloomy sky. This black torrent of countless individuals forms a vast, chaotic, ever-shifting vortex in the sky, starkly contrasting against the leaden cloud backdrop. Their silhouettes spiral upward, shattering the desolate silence and creating an atmosphere of tension filled with foreboding and primal power.", + "seed": 9871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_278_seed_9871_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031092833168804644, + "vlm_logprob": -0.031092833168804644, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.031092833168804644, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04135613888502121 + ], + "opposite_logprob_tokens": [ + -4.041356086730957 + ], + "logprob_reward": -0.04135613888502121, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04135613888502121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7517703771591187, + 0.03753127530217171, + 0.012001685798168182, + 0.026416834443807602, + 0.06144815683364868, + 0.00917330663651228, + 0.025126812979578972, + 0.035729117691516876, + 0.022577574476599693, + 0.01822482794523239 + ], + "credit_map": [ + 0.03646640479564667, + 0.27533233165740967, + 0.011301910504698753, + 0.010603824630379677, + 0.33288612961769104, + 0.05817608907818794, + 0.015954365953803062, + 0.011049359105527401, + 0.0024258093908429146, + 0.009657392278313637, + 0.0036140086594969034, + 0.004709491040557623, + 0.003323083743453026, + 0.007879313081502914, + 0.003331102430820465, + 0.0025910786353051662, + 0.0009769756579771638, + 0.001336755696684122, + 0.0019032497657462955, + 0.0016902971547096968, + 0.0009720207308419049, + 0.0012248145649209619, + 0.0019342289306223392, + 0.0019633425399661064, + 0.0014561227289959788, + 0.0022354971151798964, + 0.00335173518396914, + 0.0034468192607164383, + 0.0029729262460023165, + 0.0022842136677354574, + 0.006713678129017353, + 0.003955841530114412, + 0.002033981028944254, + 0.017128385603427887, + 0.003917227033525705, + 0.002286310773342848, + 0.02364841103553772, + 0.007020448334515095, + 0.003002642188221216, + 0.002410748740658164, + 0.0009590932750143111, + 0.0014505364233627915, + 0.0007117601344361901, + 0.0009103592019528151, + 0.0017157925758510828, + 0.0015480954898521304, + 0.0008278095629066229, + 0.0010498588671907783, + 0.004217628389596939, + 0.00395551510155201, + 0.0025377406273037195, + 0.00260176882147789, + 0.004273280035704374, + 0.0031143820378929377, + 0.002351120114326477, + 0.0020753771532326937, + 0.006913169752806425, + 0.007598270196467638, + 0.0032461502123624086, + 0.0018934670370072126, + 0.0032207334879785776, + 0.007698178756982088, + 0.0024038641713559628, + 0.002755285706371069, + 0.003319274401292205, + 0.002079569036141038, + 0.002913333475589752, + 0.0024496973492205143, + 0.0029786585364490747, + 0.0035034180618822575, + 0.0022794438991695642, + 0.003054181346669793, + 0.0024574147537350655, + 0.0024944052565842867, + 0.0028539737686514854, + 0.0012439873535186052, + 0.0023933248594403267, + 0.004154786467552185, + 0.0012909761862829328, + 0.0013359589502215385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04135613888502121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906116336584091 + ], + "opposite_logprob_tokens": [ + -2.4640612602233887 + ], + "logprob_reward": -0.08906116336584091, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906116336584091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5701273083686829, + 0.03883018717169762, + 0.0629517212510109, + 0.013883735984563828, + 0.12311071902513504, + 0.10881970822811127, + 0.021339882165193558, + 0.01822226494550705, + 0.012443888001143932, + 0.030270563438534737 + ], + "credit_map": [ + 0.021138334646821022, + 0.20731447637081146, + 0.005779599770903587, + 0.008020198903977871, + 0.25783905386924744, + 0.04717901721596718, + 0.01397381816059351, + 0.00888282060623169, + 0.0028524899389594793, + 0.009849301539361477, + 0.004096806515008211, + 0.0038156346417963505, + 0.00684068538248539, + 0.0040916865691542625, + 0.003972338512539864, + 0.0033112510573118925, + 0.004894857294857502, + 0.005351597443223, + 0.01097089797258377, + 0.009157704189419746, + 0.004203014075756073, + 0.005493887700140476, + 0.010798812843859196, + 0.012080959975719452, + 0.0013930164277553558, + 0.0023531168699264526, + 0.0011464926647022367, + 0.0015682431403547525, + 0.0034344561863690615, + 0.0018226569518446922, + 0.001219366560690105, + 0.000946387997828424, + 0.0029825367964804173, + 0.047406286001205444, + 0.0023137617390602827, + 0.001945903175510466, + 0.05437532812356949, + 0.0093661118298769, + 0.002957231132313609, + 0.0017635764088481665, + 0.015603737905621529, + 0.015538555569946766, + 0.011406398378312588, + 0.013759871944785118, + 0.01557237934321165, + 0.017049338668584824, + 0.009923129342496395, + 0.009966308251023293, + 0.0034566656686365604, + 0.002710912609472871, + 0.001882197568193078, + 0.0022665534634143114, + 0.0034312063362449408, + 0.002559542888775468, + 0.002970897825434804, + 0.0020619083661586046, + 0.0030260630883276463, + 0.0035343808121979237, + 0.001750992494635284, + 0.0009664631215855479, + 0.0020521539263427258, + 0.004410213325172663, + 0.001206098822876811, + 0.001275899587199092, + 0.001553995767608285, + 0.001518790377303958, + 0.0012088390067219734, + 0.0008849420119076967, + 0.001937410095706582, + 0.0029175940435379744, + 0.0011023671831935644, + 0.0013199503300711513, + 0.004549029748886824, + 0.004823077470064163, + 0.004666815511882305, + 0.0018880501156672835, + 0.003427583258599043, + 0.007263133302330971, + 0.0018235560273751616, + 0.0018293183529749513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.08906116336584091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787900373339653 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009787900373339653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787900373339653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5493336915969849, + 0.01585153117775917, + 0.019184518605470657, + 0.014086990617215633, + 0.09014885127544403, + 0.03791387006640434, + 0.04123777896165848, + 0.16063658893108368, + 0.0368882492184639, + 0.03471789509057999 + ], + "credit_map": [ + 0.023269928991794586, + 0.2015443742275238, + 0.006328873336315155, + 0.006431079003959894, + 0.24796989560127258, + 0.04397372528910637, + 0.012172737158834934, + 0.007643145974725485, + 0.001665766816586256, + 0.0020299365278333426, + 0.0015048801433295012, + 0.0011032624170184135, + 0.004812438040971756, + 0.002947647823020816, + 0.0009127101511694491, + 0.0008748912950977683, + 0.001507285051047802, + 0.00275668827816844, + 0.0021433765068650246, + 0.0036337089259177446, + 0.0018208721885457635, + 0.0018615778535604477, + 0.0026780811604112387, + 0.002782929688692093, + 0.0010346604976803064, + 0.0010352306999266148, + 0.0018504323670640588, + 0.0017728463280946016, + 0.0012589802499860525, + 0.0011451607570052147, + 0.003364805830642581, + 0.0026248767971992493, + 0.0032931554596871138, + 0.02594095841050148, + 0.0038795252330601215, + 0.0030042959842830896, + 0.03364140912890434, + 0.008507857099175453, + 0.004150059074163437, + 0.0077315992675721645, + 0.004190731793642044, + 0.004691538866609335, + 0.003177554812282324, + 0.002937902696430683, + 0.008606046438217163, + 0.008331161923706532, + 0.002437856048345566, + 0.0035410828422755003, + 0.005577220115810633, + 0.004925180226564407, + 0.0038771694526076317, + 0.004127503838390112, + 0.008302460424602032, + 0.004823522642254829, + 0.005091418512165546, + 0.004513306077569723, + 0.058292850852012634, + 0.05661175772547722, + 0.005280655808746815, + 0.0037435393314808607, + 0.009482719004154205, + 0.017417555674910545, + 0.004564956296235323, + 0.005242573097348213, + 0.006687341723591089, + 0.004627597518265247, + 0.003367627738043666, + 0.0032113504130393267, + 0.005527757573872805, + 0.005720307119190693, + 0.0036557100247591734, + 0.004090560134500265, + 0.0038507336284965277, + 0.004803448915481567, + 0.0041086324490606785, + 0.003232663031667471, + 0.007507947739213705, + 0.005564297549426556, + 0.0034120522905141115, + 0.002238123444840312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6954662799835205, + 0.01678829826414585, + 0.016339827328920364, + 0.012967633083462715, + 0.04404692351818085, + 0.014958064071834087, + 0.023963235318660736, + 0.08346675336360931, + 0.06409704685211182, + 0.027905941009521484 + ], + "credit_map": [ + 0.041319966316223145, + 0.25196924805641174, + 0.012477555312216282, + 0.011718735098838806, + 0.2965909242630005, + 0.049808043986558914, + 0.01790684461593628, + 0.013674951158463955, + 0.002230120589956641, + 0.0025240194518119097, + 0.001682429458014667, + 0.002070928458124399, + 0.0025007918011397123, + 0.0018729304429143667, + 0.0021516545675694942, + 0.0017554236110299826, + 0.0016018326859921217, + 0.003011250402778387, + 0.001836125971749425, + 0.00238661072216928, + 0.0017686403589323163, + 0.0017197146080434322, + 0.00187565793748945, + 0.0021399951074272394, + 0.0009080705931410193, + 0.0015307473950088024, + 0.0012175410520285368, + 0.001529942499473691, + 0.0031482011545449495, + 0.0018133297562599182, + 0.0015503907343372703, + 0.0012694093165919185, + 0.002282722620293498, + 0.014642218127846718, + 0.0020966650918126106, + 0.0017971651395782828, + 0.014466492459177971, + 0.0033176708966493607, + 0.002177364891394973, + 0.003266627434641123, + 0.0014985260786488652, + 0.0023706930223852396, + 0.0014207731001079082, + 0.001694553648121655, + 0.0024358138907700777, + 0.002785017713904381, + 0.0014465510612353683, + 0.0013061354402452707, + 0.0052707744762301445, + 0.0038288147188723087, + 0.001745553221553564, + 0.0026185945607721806, + 0.003231349168345332, + 0.00282334559597075, + 0.002376329619437456, + 0.002068473491817713, + 0.021508052945137024, + 0.018899202346801758, + 0.008866644464433193, + 0.007302179932594299, + 0.005901629570871592, + 0.008708854205906391, + 0.006633251905441284, + 0.005646936129778624, + 0.009869667701423168, + 0.0074478331953287125, + 0.006045604590326548, + 0.007601167540997267, + 0.009344145655632019, + 0.0077786012552678585, + 0.008179274387657642, + 0.007830754853785038, + 0.004266894888132811, + 0.0038558878004550934, + 0.002933795563876629, + 0.001611087704077363, + 0.004767412319779396, + 0.006698863580822945, + 0.002014699624851346, + 0.0017572984797880054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.007629304192960262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7613868713378906, + 0.052905987948179245, + 0.00900877546519041, + 0.008062662556767464, + 0.06779114156961441, + 0.02411348931491375, + 0.012186435982584953, + 0.042601775377988815, + 0.011198240332305431, + 0.010744623839855194 + ], + "credit_map": [ + 0.02804262563586235, + 0.2799796164035797, + 0.007948830723762512, + 0.007084745913743973, + 0.34806573390960693, + 0.0626508817076683, + 0.017020314931869507, + 0.010594146326184273, + 0.00494574336335063, + 0.006849510129541159, + 0.003325896104797721, + 0.0035139501560479403, + 0.014545292593538761, + 0.013085485436022282, + 0.0036299768835306168, + 0.0030101281590759754, + 0.0008458462543785572, + 0.0007909529958851635, + 0.001125034294091165, + 0.001550032407976687, + 0.0010472420835867524, + 0.0013970930594950914, + 0.0012849824270233512, + 0.0009675921173766255, + 0.000422386423451826, + 0.0013656083028763533, + 0.0008800613577477634, + 0.0008052074117586017, + 0.001470678485929966, + 0.0011210477678105235, + 0.0011286293156445026, + 0.0008690436370670795, + 0.0019702990539371967, + 0.028675194829702377, + 0.0012455922551453114, + 0.001106869662180543, + 0.027437932789325714, + 0.0044000702910125256, + 0.0016291063511744142, + 0.0013260783161967993, + 0.0034243448171764612, + 0.002970366971567273, + 0.0021801304537802935, + 0.0022145879920572042, + 0.004213697277009487, + 0.0047171032056212425, + 0.0022492497228085995, + 0.002144008642062545, + 0.0017250851960852742, + 0.001800879486836493, + 0.0012515063863247633, + 0.0010993932373821735, + 0.0020558240357786417, + 0.0015029233181849122, + 0.00160332175437361, + 0.0011475023347884417, + 0.015362361446022987, + 0.014514820650219917, + 0.002070913789793849, + 0.0018404577858746052, + 0.0021816024091094732, + 0.0035182791762053967, + 0.0016011359402909875, + 0.0015122059267014265, + 0.0016040044138208032, + 0.0014831123407930136, + 0.0010432528797537088, + 0.0008452003239654005, + 0.0017089807661250234, + 0.0024822514969855547, + 0.0009024114697240293, + 0.001129027339629829, + 0.0015756976790726185, + 0.0017429516883566976, + 0.001706455834209919, + 0.0009641227661632001, + 0.0013369293883442879, + 0.0015110376989468932, + 0.0010593781480565667, + 0.0008480508695356548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.007629659026861191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a low angle, the camera tilts upwards at a slow, steady pace, tracking a sudden commotion. A vast expanse of wilderness, the ground covered with withered, dry grass, dotted with a few twisted, leafless trees. The sky is shrouded in thick, swirling leaden clouds, and the faint, diffuse light casts an oppressive, monochromatic hue. Hundreds, perhaps thousands, of pitch-black crows, their feathers gleaming a glossy blue-purple in the gloom, roost on the ground and on dead branches. As if startled by an invisible, resounding roar, they suddenly flap their wings, emitting a chaotic, piercing cawing noise, and like an exploding black wave, they surge into the gloomy sky. This black torrent of countless individuals forms a vast, chaotic, ever-shifting vortex in the sky, starkly contrasting against the leaden cloud backdrop. Their silhouettes spiral upward, shattering the desolate silence and creating an atmosphere of tension filled with foreboding and primal power.", + "seed": 9872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_278_seed_9872_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016734612919390202, + "vlm_logprob": -0.016734612919390202, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.016734612919390202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04109474644064903 + ], + "opposite_logprob_tokens": [ + -3.9160947799682617 + ], + "logprob_reward": -0.04109474644064903, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04109474644064903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14617694914340973, + 0.09051311016082764, + 0.12806294858455658, + 0.05239708349108696, + 0.054598551243543625, + 0.06916926056146622, + 0.15998625755310059, + 0.11835819482803345, + 0.11278694868087769, + 0.06795063614845276 + ], + "credit_map": [ + 0.021756278350949287, + 0.03175957873463631, + 0.014783159829676151, + 0.01923825964331627, + 0.020393120124936104, + 0.014628643169999123, + 0.012401665560901165, + 0.01121625117957592, + 0.008016012609004974, + 0.008739166893064976, + 0.008798603899776936, + 0.012985729612410069, + 0.012115225195884705, + 0.011756950058043003, + 0.017247792333364487, + 0.010853623040020466, + 0.021259594708681107, + 0.012546501122415066, + 0.017019052058458328, + 0.012756397016346455, + 0.009853394702076912, + 0.010763274505734444, + 0.019791485741734505, + 0.02407325990498066, + 0.005694159772247076, + 0.0039015512447804213, + 0.008679143153131008, + 0.00800166092813015, + 0.003713905345648527, + 0.0032108307350426912, + 0.010590823367238045, + 0.008605008013546467, + 0.011895515955984592, + 0.007507863454520702, + 0.006522957235574722, + 0.005228795111179352, + 0.005931858438998461, + 0.0051885065622627735, + 0.007241363637149334, + 0.005081693176180124, + 0.007744461297988892, + 0.006665689870715141, + 0.010902655310928822, + 0.008122282102704048, + 0.009249613620340824, + 0.006305404007434845, + 0.010991603136062622, + 0.009187547490000725, + 0.01900419220328331, + 0.01646459475159645, + 0.01864686980843544, + 0.015132035128772259, + 0.025628438219428062, + 0.022867267951369286, + 0.018202366307377815, + 0.024040497839450836, + 0.020671691745519638, + 0.017734592780470848, + 0.01691116951406002, + 0.013699681498110294, + 0.00911694671958685, + 0.008041941560804844, + 0.012883101589977741, + 0.019299065694212914, + 0.014250053092837334, + 0.01750200428068638, + 0.02274562045931816, + 0.021319633349776268, + 0.010423569940030575, + 0.009034369140863419, + 0.008920135907828808, + 0.008591566234827042, + 0.013977272436022758, + 0.012031514197587967, + 0.008352243341505527, + 0.010912403464317322, + 0.004552820697426796, + 0.004471823573112488, + 0.00721427146345377, + 0.006438293494284153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.04109474644064903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096433736383915 + ], + "opposite_logprob_tokens": [ + -4.511096477508545 + ], + "logprob_reward": -0.011096433736383915, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096433736383915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2046799659729004, + 0.06831379234790802, + 0.15728440880775452, + 0.14595863223075867, + 0.05115630477666855, + 0.04391149803996086, + 0.07882190495729446, + 0.15946343541145325, + 0.0421803742647171, + 0.048229604959487915 + ], + "credit_map": [ + 0.02981354109942913, + 0.043785952031612396, + 0.016628611832857132, + 0.022631891071796417, + 0.027022074908018112, + 0.02782883495092392, + 0.0208094771951437, + 0.016159599646925926, + 0.005963452626019716, + 0.005095287691801786, + 0.006128254812210798, + 0.007795348297804594, + 0.006037835497409105, + 0.00763064855709672, + 0.0178143959492445, + 0.01184857077896595, + 0.026536526158452034, + 0.017632363364100456, + 0.022648489102721214, + 0.016569657251238823, + 0.016643475741147995, + 0.014633170329034328, + 0.022191016003489494, + 0.020429719239473343, + 0.018546193838119507, + 0.023515598848462105, + 0.011380604468286037, + 0.013076985254883766, + 0.009792671538889408, + 0.00916479155421257, + 0.02405640110373497, + 0.0364253856241703, + 0.0076960911974310875, + 0.006963878870010376, + 0.005582922603935003, + 0.0038746031932532787, + 0.0036429925821721554, + 0.0028677305672317743, + 0.009378550574183464, + 0.011149533092975616, + 0.0064077796414494514, + 0.00833717081695795, + 0.005622540134936571, + 0.005908631719648838, + 0.0034075756557285786, + 0.0029530273750424385, + 0.005464806221425533, + 0.005809968337416649, + 0.009869029745459557, + 0.010355433449149132, + 0.011256121098995209, + 0.014016249217092991, + 0.006859636399894953, + 0.005362918600440025, + 0.013156934641301632, + 0.007945586927235126, + 0.03246025741100311, + 0.018821021541953087, + 0.026531625539064407, + 0.03253705054521561, + 0.011676189489662647, + 0.011899850331246853, + 0.014701594598591328, + 0.010835847817361355, + 0.004045527428388596, + 0.004475811962038279, + 0.010001782327890396, + 0.007275447715073824, + 0.006025548093020916, + 0.003127477364614606, + 0.0036797821521759033, + 0.0035489960573613644, + 0.011062118224799633, + 0.00929354876279831, + 0.00554449949413538, + 0.007828159257769585, + 0.0027741161175072193, + 0.003369478974491358, + 0.003736370475962758, + 0.004621314350515604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.011096433736383915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640522137284279 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640522137284279, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640522137284279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16735154390335083, + 0.16518044471740723, + 0.07573133707046509, + 0.040685493499040604, + 0.09767749160528183, + 0.12078049778938293, + 0.0941968709230423, + 0.08201954513788223, + 0.10440001636743546, + 0.05197674781084061 + ], + "credit_map": [ + 0.020010437816381454, + 0.03796873241662979, + 0.010274231433868408, + 0.016774814575910568, + 0.027701348066329956, + 0.021967310458421707, + 0.02166905626654625, + 0.01098562777042389, + 0.01276800874620676, + 0.011411682702600956, + 0.017825953662395477, + 0.0232402253895998, + 0.013373720459640026, + 0.019205501303076744, + 0.04134877026081085, + 0.026006586849689484, + 0.01211098674684763, + 0.007036480121314526, + 0.010745055042207241, + 0.008272327482700348, + 0.006179722025990486, + 0.007815233431756496, + 0.009877157397568226, + 0.013694378547370434, + 0.009032085537910461, + 0.005429517477750778, + 0.004296864848583937, + 0.0043280115351080894, + 0.003016360104084015, + 0.0025569091085344553, + 0.006492300424724817, + 0.005533446092158556, + 0.013337722979485989, + 0.009286877699196339, + 0.015392881818115711, + 0.013960063457489014, + 0.008715560659766197, + 0.00860634259879589, + 0.01369816716760397, + 0.014679886400699615, + 0.013397731818258762, + 0.012058964930474758, + 0.011788692325353622, + 0.01823355443775654, + 0.022748151794075966, + 0.01217124704271555, + 0.01616887003183365, + 0.01421329379081726, + 0.016677331179380417, + 0.010861586779356003, + 0.007835406810045242, + 0.008507993072271347, + 0.011974467895925045, + 0.01345313061028719, + 0.016387594863772392, + 0.00849936157464981, + 0.009995050728321075, + 0.009219028986990452, + 0.01187037955969572, + 0.01500380877405405, + 0.00724215479567647, + 0.00912029854953289, + 0.01211366057395935, + 0.007455165032297373, + 0.01649704948067665, + 0.02340981364250183, + 0.009832949377596378, + 0.011895647272467613, + 0.011600105091929436, + 0.012302575632929802, + 0.008731127716600895, + 0.010130750946700573, + 0.009557105600833893, + 0.009301184676587582, + 0.006874088663607836, + 0.005523952189832926, + 0.004672663286328316, + 0.005347569938749075, + 0.005238104145973921, + 0.005462081637233496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008640522137284279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296156793832779, + 0.0892195776104927, + 0.07525064051151276, + 0.054082151502370834, + 0.08477997779846191, + 0.06293725967407227, + 0.17926231026649475, + 0.09507060050964355, + 0.1147356629371643, + 0.11504608392715454 + ], + "credit_map": [ + 0.017425818368792534, + 0.02265840768814087, + 0.011740747839212418, + 0.012539515271782875, + 0.022865522652864456, + 0.01772894710302353, + 0.013582944869995117, + 0.011073775589466095, + 0.009645742364227772, + 0.007634333800524473, + 0.006846196483820677, + 0.00957572367042303, + 0.007172232028096914, + 0.013445853255689144, + 0.021512426435947418, + 0.01338706910610199, + 0.008923860266804695, + 0.006814624648541212, + 0.009483539499342442, + 0.008644486777484417, + 0.00519587891176343, + 0.007246978115290403, + 0.011323131620883942, + 0.017618143931031227, + 0.011319143697619438, + 0.010417548008263111, + 0.005403328221291304, + 0.0050154379568994045, + 0.003287726314738393, + 0.003058439353480935, + 0.00820031575858593, + 0.0073802112601697445, + 0.019030103459954262, + 0.0090679582208395, + 0.004096880089491606, + 0.005929281935095787, + 0.0037725488655269146, + 0.004689904861152172, + 0.01429247111082077, + 0.02390083111822605, + 0.007957887835800648, + 0.007853020913898945, + 0.007826696150004864, + 0.0070047881454229355, + 0.012503150850534439, + 0.009611348621547222, + 0.005438932217657566, + 0.004741437267512083, + 0.02398635819554329, + 0.019204920157790184, + 0.014469536021351814, + 0.01567510887980461, + 0.026163650676608086, + 0.03330926597118378, + 0.02362525276839733, + 0.022828221321105957, + 0.014798597432672977, + 0.010021023452281952, + 0.01820530742406845, + 0.016907891258597374, + 0.008463213220238686, + 0.010143542662262917, + 0.00839550793170929, + 0.008135519921779633, + 0.01495328824967146, + 0.02239726111292839, + 0.017237359657883644, + 0.018389875069260597, + 0.012551605701446533, + 0.011685458943247795, + 0.0101480633020401, + 0.0073727574199438095, + 0.014898664318025112, + 0.01761299930512905, + 0.016850285232067108, + 0.01320648193359375, + 0.013246393762528896, + 0.014943874441087246, + 0.01169878151267767, + 0.012588605284690857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.008643358945846558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198003336787224 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014198003336787224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198003336787224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3374672830104828, + 0.07856050133705139, + 0.18036749958992004, + 0.02626108005642891, + 0.037501074373722076, + 0.043817680329084396, + 0.02061007358133793, + 0.0645950436592102, + 0.04677832871675491, + 0.1640414446592331 + ], + "credit_map": [ + 0.04108693078160286, + 0.053367238491773605, + 0.022854968905448914, + 0.02881171740591526, + 0.07306625694036484, + 0.05083225667476654, + 0.03919564560055733, + 0.02825227752327919, + 0.008421625941991806, + 0.00970706157386303, + 0.006233722437173128, + 0.009229637682437897, + 0.012871455401182175, + 0.011785137467086315, + 0.012995792552828789, + 0.007316068280488253, + 0.02629789151251316, + 0.02444152534008026, + 0.02797502465546131, + 0.018038274720311165, + 0.016084864735603333, + 0.01966424286365509, + 0.024794364348053932, + 0.023071300238370895, + 0.0057249353267252445, + 0.004751544911414385, + 0.002980451565235853, + 0.002639448968693614, + 0.0021265975665301085, + 0.0016194636700674891, + 0.0035587414167821407, + 0.0028598972130566835, + 0.0033450950868427753, + 0.0047290935181081295, + 0.0043509285897016525, + 0.0041190446354448795, + 0.0024171341210603714, + 0.002412657020613551, + 0.006802681367844343, + 0.009324437938630581, + 0.005260583478957415, + 0.004329435992985964, + 0.00430554524064064, + 0.007162777706980705, + 0.0045821769163012505, + 0.003936225548386574, + 0.008279162459075451, + 0.005961777176707983, + 0.002906931098550558, + 0.0016755044925957918, + 0.003252338385209441, + 0.0024360078386962414, + 0.00257134810090065, + 0.002645008498802781, + 0.00287953345105052, + 0.002243401948362589, + 0.015655208379030228, + 0.009641094133257866, + 0.010591035708785057, + 0.008997068740427494, + 0.00393848679959774, + 0.0036496352404356003, + 0.005917261820286512, + 0.00620524724945426, + 0.005117130931466818, + 0.005649207625538111, + 0.009244151413440704, + 0.008900950662791729, + 0.004982450976967812, + 0.004359255079180002, + 0.0038432693108916283, + 0.004681918304413557, + 0.03685176372528076, + 0.03647033870220184, + 0.02103513479232788, + 0.02551260031759739, + 0.009519997052848339, + 0.010346812196075916, + 0.010803744196891785, + 0.01350104995071888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.014198003336787224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a low angle, the camera tilts upwards at a slow, steady pace, tracking a sudden commotion. A vast expanse of wilderness, the ground covered with withered, dry grass, dotted with a few twisted, leafless trees. The sky is shrouded in thick, swirling leaden clouds, and the faint, diffuse light casts an oppressive, monochromatic hue. Hundreds, perhaps thousands, of pitch-black crows, their feathers gleaming a glossy blue-purple in the gloom, roost on the ground and on dead branches. As if startled by an invisible, resounding roar, they suddenly flap their wings, emitting a chaotic, piercing cawing noise, and like an exploding black wave, they surge into the gloomy sky. This black torrent of countless individuals forms a vast, chaotic, ever-shifting vortex in the sky, starkly contrasting against the leaden cloud backdrop. Their silhouettes spiral upward, shattering the desolate silence and creating an atmosphere of tension filled with foreboding and primal power.", + "seed": 9873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_278_seed_9873_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01608677553012967, + "vlm_logprob": -0.01608677553012967, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01608677553012967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04376527667045593 + ], + "opposite_logprob_tokens": [ + -3.9187653064727783 + ], + "logprob_reward": -0.04376527667045593, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04376527667045593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793402135372162, + 0.11166466027498245, + 0.09411439299583435, + 0.14033766090869904, + 0.055004943162202835, + 0.0920868068933487, + 0.0622221976518631, + 0.08646587282419205, + 0.0522235669195652, + 0.1265396922826767 + ], + "credit_map": [ + 0.017108120024204254, + 0.02161603607237339, + 0.013521677814424038, + 0.026442768052220345, + 0.03575998917222023, + 0.022906558588147163, + 0.026387592777609825, + 0.015597494319081306, + 0.020288385450839996, + 0.026181895285844803, + 0.008162385784089565, + 0.01164296641945839, + 0.01460101455450058, + 0.011935777962207794, + 0.01365441270172596, + 0.005197825841605663, + 0.01035529188811779, + 0.011821620166301727, + 0.0069900997914373875, + 0.011441415175795555, + 0.014711962081491947, + 0.013325021602213383, + 0.01618175394833088, + 0.009287218563258648, + 0.022859616205096245, + 0.025605501607060432, + 0.006830333266407251, + 0.00994709599763155, + 0.03136862441897392, + 0.021298794075846672, + 0.014612014405429363, + 0.007815679535269737, + 0.006363343447446823, + 0.004427880514413118, + 0.005030829459428787, + 0.006128971930593252, + 0.00577486352995038, + 0.009363753721117973, + 0.008953179232776165, + 0.008962118998169899, + 0.011272828094661236, + 0.01910148561000824, + 0.007599132601171732, + 0.009408343583345413, + 0.011933060362935066, + 0.015742609277367592, + 0.010015062987804413, + 0.007014281582087278, + 0.007719260174781084, + 0.00794364232569933, + 0.009417214430868626, + 0.00619961554184556, + 0.009112242609262466, + 0.008605255745351315, + 0.007179886102676392, + 0.0060450793243944645, + 0.013542582280933857, + 0.010561305098235607, + 0.017697270959615707, + 0.00991691742092371, + 0.0067666200920939445, + 0.010536056943237782, + 0.00876594427973032, + 0.008679174818098545, + 0.008456903509795666, + 0.005427909549325705, + 0.009209941141307354, + 0.005605361890047789, + 0.006633819546550512, + 0.006905820686370134, + 0.004135094583034515, + 0.005848721135407686, + 0.027790138497948647, + 0.010700815357267857, + 0.026387104764580727, + 0.00959017500281334, + 0.005507460329681635, + 0.007475708145648241, + 0.024768445640802383, + 0.014319838024675846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.04376527667045593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059118263423443 + ], + "opposite_logprob_tokens": [ + -3.8955912590026855 + ], + "logprob_reward": -0.02059118263423443, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059118263423443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6317563652992249, + 0.10796201229095459, + 0.04368928447365761, + 0.06954753398895264, + 0.025259045884013176, + 0.016336046159267426, + 0.025699350982904434, + 0.04292304813861847, + 0.02369639091193676, + 0.013130966573953629 + ], + "credit_map": [ + 0.06183120235800743, + 0.13050496578216553, + 0.037035830318927765, + 0.09037419408559799, + 0.1364540308713913, + 0.06679179519414902, + 0.0699976235628128, + 0.038766760379076004, + 0.012055795639753342, + 0.018108606338500977, + 0.005588800646364689, + 0.00616608839482069, + 0.03387627750635147, + 0.015623928979039192, + 0.010741385631263256, + 0.00580113148316741, + 0.008102613501250744, + 0.004285994917154312, + 0.003356729168444872, + 0.0035002732183784246, + 0.0073328218422830105, + 0.005789745599031448, + 0.006709131877869368, + 0.004611970856785774, + 0.008192762732505798, + 0.01871914602816105, + 0.0038389319088310003, + 0.0026825182139873505, + 0.021733198314905167, + 0.007092813029885292, + 0.004357344936579466, + 0.0029308164957910776, + 0.00601368909701705, + 0.0030905408784747124, + 0.001783038955181837, + 0.0024671326391398907, + 0.0036509002093225718, + 0.0039854696951806545, + 0.001957420725375414, + 0.0023108527529984713, + 0.002426575869321823, + 0.0023698334116488695, + 0.0014687859220430255, + 0.0015577793819829822, + 0.0019311474170535803, + 0.001871646149083972, + 0.0032719688024371862, + 0.001438308972865343, + 0.0026811212301254272, + 0.004236247390508652, + 0.0029643713496625423, + 0.0021902371663600206, + 0.003637858899310231, + 0.004559758584946394, + 0.002401108620688319, + 0.0030286468099802732, + 0.00567705137655139, + 0.0039046143647283316, + 0.009664826095104218, + 0.005532294977456331, + 0.003772263415157795, + 0.005271148402243853, + 0.0036086023319512606, + 0.005492243450134993, + 0.0031658203806728125, + 0.00334177422337234, + 0.0031520964112132788, + 0.002318244194611907, + 0.004280898720026016, + 0.0025908597745001316, + 0.0018419679254293442, + 0.0030047285836189985, + 0.002880855230614543, + 0.001976186176761985, + 0.002386292675510049, + 0.0014003715477883816, + 0.0007737223641015589, + 0.0008451117319054902, + 0.0011191755766049027, + 0.0017492512706667185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02059118263423443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251422990113497 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251422990113497, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251422990113497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16303642094135284, + 0.1385718286037445, + 0.07483265548944473, + 0.1364699900150299, + 0.08977355062961578, + 0.09162743389606476, + 0.05178934708237648, + 0.12541000545024872, + 0.07123179733753204, + 0.05725693702697754 + ], + "credit_map": [ + 0.017138265073299408, + 0.024952346459031105, + 0.012566672638058662, + 0.01707208901643753, + 0.03642359375953674, + 0.023130318149924278, + 0.020820388570427895, + 0.010932748205959797, + 0.017005503177642822, + 0.029593372717499733, + 0.006922876462340355, + 0.00663589034229517, + 0.046052053570747375, + 0.020242605358362198, + 0.008302871137857437, + 0.003816652577370405, + 0.016030633822083473, + 0.01395613793283701, + 0.0037482876796275377, + 0.0038094930350780487, + 0.01952233538031578, + 0.010866396129131317, + 0.0036910329945385456, + 0.003208338748663664, + 0.02736128866672516, + 0.012526472099125385, + 0.011044656857848167, + 0.014608447439968586, + 0.02045416459441185, + 0.024793894961476326, + 0.014840681105852127, + 0.010840382426977158, + 0.018683362752199173, + 0.011292822659015656, + 0.007352136075496674, + 0.008484257385134697, + 0.013291021808981895, + 0.011938974261283875, + 0.006898066494613886, + 0.011832904070615768, + 0.009049546904861927, + 0.014451487921178341, + 0.008331089280545712, + 0.010470095090568066, + 0.009675711393356323, + 0.015570935793220997, + 0.015177056193351746, + 0.008901511318981647, + 0.00802539475262165, + 0.0073477295227348804, + 0.006936131976544857, + 0.00462111784145236, + 0.007335749454796314, + 0.008574147708714008, + 0.0045335739850997925, + 0.004415501840412617, + 0.016404803842306137, + 0.017685115337371826, + 0.021517356857657433, + 0.014415720477700233, + 0.014373286627233028, + 0.020589830353856087, + 0.01100009772926569, + 0.00942379329353571, + 0.008940731175243855, + 0.00890908669680357, + 0.007905500940978527, + 0.00812012329697609, + 0.006681653670966625, + 0.010165493935346603, + 0.006651146803051233, + 0.013858062215149403, + 0.007819396443665028, + 0.009099816903471947, + 0.005810015834867954, + 0.006432826165109873, + 0.005347928032279015, + 0.005016359966248274, + 0.011533486656844616, + 0.006197106558829546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005251422990113497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734649650752544 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734649650752544, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734649650752544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13762958347797394, + 0.2214357703924179, + 0.07549925148487091, + 0.1158042699098587, + 0.05789410322904587, + 0.029959848150610924, + 0.06964829564094543, + 0.10159356892108917, + 0.07397785782814026, + 0.11655743420124054 + ], + "credit_map": [ + 0.01624344103038311, + 0.03211205452680588, + 0.010126917622983456, + 0.01705159991979599, + 0.027873024344444275, + 0.013200651854276657, + 0.013653220608830452, + 0.007368660997599363, + 0.019732965156435966, + 0.046046093106269836, + 0.008590267971158028, + 0.012190858833491802, + 0.07755346596240997, + 0.029671190306544304, + 0.020962825044989586, + 0.0066881077364087105, + 0.019937090575695038, + 0.014183328486979008, + 0.0056169237941503525, + 0.004895597696304321, + 0.009147224016487598, + 0.007092352956533432, + 0.00863806065171957, + 0.005988666787743568, + 0.0072941044345498085, + 0.02936125546693802, + 0.006014589220285416, + 0.005453602876514196, + 0.03685200586915016, + 0.014895813539624214, + 0.010522226803004742, + 0.005410674028098583, + 0.01417112909257412, + 0.006420116405934095, + 0.0045622666366398335, + 0.004527552053332329, + 0.009550116956233978, + 0.0076356022618710995, + 0.005582732148468494, + 0.005444588139653206, + 0.004267632495611906, + 0.005242169369012117, + 0.003882104530930519, + 0.0036971981171518564, + 0.003927319310605526, + 0.0032140372786670923, + 0.003477102145552635, + 0.0022522832732647657, + 0.00806748028844595, + 0.012328742071986198, + 0.007770055904984474, + 0.0072251115925610065, + 0.010189633816480637, + 0.012853069230914116, + 0.005564755294471979, + 0.005649448372423649, + 0.01388128288090229, + 0.013058137148618698, + 0.023498300462961197, + 0.013128320686519146, + 0.009200355969369411, + 0.011556376703083515, + 0.00784655474126339, + 0.009424237534403801, + 0.00900230836123228, + 0.014491125009953976, + 0.008419794030487537, + 0.010430867783725262, + 0.0058298613876104355, + 0.01167332474142313, + 0.006392962299287319, + 0.007737616077065468, + 0.042095329612493515, + 0.0058066388592123985, + 0.03841394558548927, + 0.007214197888970375, + 0.0029026109259575605, + 0.004694716073572636, + 0.010226240381598473, + 0.0052037606947124004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006734649650752544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091345705091953 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091345705091953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091345705091953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21139343082904816, + 0.16047781705856323, + 0.11823367327451706, + 0.27393361926078796, + 0.01530886348336935, + 0.025438692420721054, + 0.02857985347509384, + 0.0201311893761158, + 0.03361539915204048, + 0.11288739740848541 + ], + "credit_map": [ + 0.019661689177155495, + 0.044404711574316025, + 0.013059848919510841, + 0.027311403304338455, + 0.044648271054029465, + 0.026937156915664673, + 0.02291862852871418, + 0.01245173066854477, + 0.012300373986363411, + 0.03139103576540947, + 0.006297004874795675, + 0.011481727473437786, + 0.054467812180519104, + 0.022008903324604034, + 0.01742706261575222, + 0.005103905685245991, + 0.014056484214961529, + 0.01996057853102684, + 0.005649876780807972, + 0.007262492086738348, + 0.032544031739234924, + 0.021596690639853477, + 0.01090281829237938, + 0.006260701920837164, + 0.03427111357450485, + 0.057462386786937714, + 0.014621260575950146, + 0.019689952954649925, + 0.06944300979375839, + 0.03885135054588318, + 0.026971012353897095, + 0.012623518705368042, + 0.0032168591860681772, + 0.0018429167103022337, + 0.00135666667483747, + 0.0006976532167755067, + 0.0022811649832874537, + 0.002749714534729719, + 0.0015139668248593807, + 0.0016499219927936792, + 0.003729849820956588, + 0.0037031129468232393, + 0.002978404052555561, + 0.002064261119812727, + 0.0037443912588059902, + 0.003435765393078327, + 0.0029670903459191322, + 0.0028158160857856274, + 0.0036289766430854797, + 0.005299043841660023, + 0.004038279876112938, + 0.0024972399696707726, + 0.004034648183733225, + 0.0049238670617341995, + 0.0018722345121204853, + 0.0022855629213154316, + 0.0030671560671180487, + 0.003072859486564994, + 0.003806644817814231, + 0.00244020763784647, + 0.0016069387784227729, + 0.0026694058906286955, + 0.001865699072368443, + 0.0016022789059206843, + 0.005465302150696516, + 0.002697373041883111, + 0.006932467222213745, + 0.0018632157007232308, + 0.006119889672845602, + 0.004267161246389151, + 0.002798350527882576, + 0.0034716424997895956, + 0.042066581547260284, + 0.006013341713696718, + 0.03807799518108368, + 0.00671934150159359, + 0.0022546800319105387, + 0.002965315245091915, + 0.009469304233789444, + 0.00532084796577692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004091345705091953 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, filmed from a low angle, the camera tilts upwards at a slow, steady pace, tracking a sudden commotion. A vast expanse of wilderness, the ground covered with withered, dry grass, dotted with a few twisted, leafless trees. The sky is shrouded in thick, swirling leaden clouds, and the faint, diffuse light casts an oppressive, monochromatic hue. Hundreds, perhaps thousands, of pitch-black crows, their feathers gleaming a glossy blue-purple in the gloom, roost on the ground and on dead branches. As if startled by an invisible, resounding roar, they suddenly flap their wings, emitting a chaotic, piercing cawing noise, and like an exploding black wave, they surge into the gloomy sky. This black torrent of countless individuals forms a vast, chaotic, ever-shifting vortex in the sky, starkly contrasting against the leaden cloud backdrop. Their silhouettes spiral upward, shattering the desolate silence and creating an atmosphere of tension filled with foreboding and primal power.", + "seed": 9874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_0/index_278_seed_9874_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023019602242857217, + "vlm_logprob": -0.023019602242857217, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.023019602242857217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031647615134716034 + ], + "opposite_logprob_tokens": [ + -3.9066476821899414 + ], + "logprob_reward": -0.031647615134716034, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.031647615134716034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13732489943504333, + 0.47135841846466064, + 0.07440493255853653, + 0.051565177738666534, + 0.02652495913207531, + 0.03544846922159195, + 0.0301648061722517, + 0.03625311702489853, + 0.03892389312386513, + 0.09803137928247452 + ], + "credit_map": [ + 0.014298245310783386, + 0.023346293717622757, + 0.009362898766994476, + 0.0089109493419528, + 0.04545656219124794, + 0.015182510018348694, + 0.011612932197749615, + 0.009154503233730793, + 0.02208428643643856, + 0.18802553415298462, + 0.00966763123869896, + 0.00958680547773838, + 0.19203415513038635, + 0.03424462303519249, + 0.008802003227174282, + 0.006913370918482542, + 0.007506208028644323, + 0.013442734256386757, + 0.009197402745485306, + 0.006831827107816935, + 0.013465606607496738, + 0.012161698192358017, + 0.00745003717020154, + 0.004349423106759787, + 0.00784088484942913, + 0.005727014504373074, + 0.007662532385438681, + 0.007003355771303177, + 0.008097860962152481, + 0.00599057087674737, + 0.004543725401163101, + 0.004699233453720808, + 0.003406010800972581, + 0.004778144881129265, + 0.00352460821159184, + 0.0032248664647340775, + 0.0030019294936209917, + 0.002937520621344447, + 0.0029693408869206905, + 0.0026825370732694864, + 0.004077494610100985, + 0.0038328133523464203, + 0.008695743046700954, + 0.0044548409059643745, + 0.0027287446428090334, + 0.003937268164008856, + 0.003819912439212203, + 0.003901652293279767, + 0.00495142163708806, + 0.006193671841174364, + 0.004339398816227913, + 0.0027018534019589424, + 0.0039021542761474848, + 0.002910548122599721, + 0.0017829322023317218, + 0.003382827853783965, + 0.006970859598368406, + 0.01137452945113182, + 0.004470191430300474, + 0.0033781733363866806, + 0.004359540529549122, + 0.00280161015689373, + 0.0012129630194976926, + 0.0016852522967383265, + 0.006258539855480194, + 0.00994829647243023, + 0.00586203346028924, + 0.003701436799019575, + 0.005195995792746544, + 0.0027963852044194937, + 0.0024763019755482674, + 0.002684908453375101, + 0.018907716497778893, + 0.018065113574266434, + 0.012251724489033222, + 0.011690440587699413, + 0.010230462066829205, + 0.010293109342455864, + 0.006630376446992159, + 0.009962433949112892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.031647615134716034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054978106170892715 + ], + "opposite_logprob_tokens": [ + -2.929978132247925 + ], + "logprob_reward": -0.054978106170892715, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054978106170892715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04120940342545509, + 0.8741567730903625, + 0.02944207563996315, + 0.02134210430085659, + 0.007886787876486778, + 0.009545691311359406, + 0.0022254555951803923, + 0.005548796616494656, + 0.003689131699502468, + 0.004953739698976278 + ], + "credit_map": [ + 0.003623218508437276, + 0.005342508666217327, + 0.0072169229388237, + 0.004658065270632505, + 0.007125547621399164, + 0.0030174795538187027, + 0.004217545967549086, + 0.006008119322359562, + 0.04063700884580612, + 0.360208123922348, + 0.008963903412222862, + 0.0156791340559721, + 0.3646768033504486, + 0.061654649674892426, + 0.012509007006883621, + 0.009828131645917892, + 0.0027585767675191164, + 0.004972476977854967, + 0.004983967170119286, + 0.0034574861638247967, + 0.003804221749305725, + 0.004182341508567333, + 0.0032061890233308077, + 0.0020768186077475548, + 0.0025469178799539804, + 0.0028601051308214664, + 0.0020981442648917437, + 0.0025836273562163115, + 0.0037756189703941345, + 0.0026007494889199734, + 0.0034187266137450933, + 0.001458215992897749, + 0.0009243727545253932, + 0.0007903210935182869, + 0.0012845539022237062, + 0.0010985132539644837, + 0.0009387676836922765, + 0.0010141081875190139, + 0.0008893644553609192, + 0.0009467870695516467, + 0.0012865934986621141, + 0.0014178010169416666, + 0.0011051716282963753, + 0.001075154053978622, + 0.0020746432710438967, + 0.0010816201101988554, + 0.0007073888555169106, + 0.000797319458797574, + 0.0003114565624855459, + 0.0003711557073984295, + 0.0002985884202644229, + 0.00019523309310898185, + 0.00023635719844605774, + 0.00024455576203763485, + 0.00022505383822135627, + 0.00034305526060052216, + 0.0009328320156782866, + 0.0012875567190349102, + 0.0006690152804367244, + 0.000482085655676201, + 0.0006704936968162656, + 0.0008548952173441648, + 0.00026048082509078085, + 0.00039143735193647444, + 0.0005525442538782954, + 0.0006195151363499463, + 0.0006336629739962518, + 0.0004886924289166927, + 0.000458561786217615, + 0.00029600426205433905, + 0.00032806902891024947, + 0.0003120821784250438, + 0.0008909015450626612, + 0.0007849677349440753, + 0.0007907373365014791, + 0.0008442850667051971, + 0.0006295769126154482, + 0.0003845901810564101, + 0.00031716839293949306, + 0.00031151293660514057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.054978106170892715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195063676685095 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.003195063676685095, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195063676685095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0734926089644432, + 0.7172783613204956, + 0.02867623046040535, + 0.023341596126556396, + 0.009572530165314674, + 0.033758360892534256, + 0.027053823694586754, + 0.03797103464603424, + 0.020174862816929817, + 0.028680510818958282 + ], + "credit_map": [ + 0.004455276299268007, + 0.01417924091219902, + 0.006421692203730345, + 0.0043955231085419655, + 0.024008970707654953, + 0.008074004203081131, + 0.0045834751799702644, + 0.0073744249530136585, + 0.03297166898846626, + 0.29465046525001526, + 0.007050951011478901, + 0.0126809636130929, + 0.30021387338638306, + 0.05129639059305191, + 0.010381036438047886, + 0.008033061400055885, + 0.003409505821764469, + 0.0064556170254945755, + 0.0033307322300970554, + 0.0031347249168902636, + 0.004557058215141296, + 0.0030186246149241924, + 0.0030418429523706436, + 0.0017281260807067156, + 0.0029509125743061304, + 0.003389619756489992, + 0.0029535051435232162, + 0.0033484059385955334, + 0.0034912617411464453, + 0.0025843342300504446, + 0.002904653549194336, + 0.0017189016798511147, + 0.0012766836443915963, + 0.0015477623092010617, + 0.0010508379200473428, + 0.0009952670661732554, + 0.0012370033655315638, + 0.0015156713780015707, + 0.001034054788760841, + 0.0009152498678304255, + 0.004570711869746447, + 0.004718815442174673, + 0.003725901711732149, + 0.00388901773840189, + 0.007094365078955889, + 0.003787811379879713, + 0.003181248204782605, + 0.0027904915623366833, + 0.0043714772909879684, + 0.0037718371022492647, + 0.003039398929104209, + 0.0026047530118376017, + 0.0045637465082108974, + 0.004136315081268549, + 0.0020205089822411537, + 0.0025457884185016155, + 0.006763891316950321, + 0.0074929906986653805, + 0.0037683802656829357, + 0.004454076290130615, + 0.003940267954021692, + 0.005666996818035841, + 0.00198628893122077, + 0.003898144932463765, + 0.002665943931788206, + 0.0034643097314983606, + 0.0034294065553694963, + 0.0033354596234858036, + 0.002273682039231062, + 0.002431244822219014, + 0.0013413183623924851, + 0.0012334984494373202, + 0.003636073786765337, + 0.0037520751357078552, + 0.005766129586845636, + 0.00451020197942853, + 0.0030747444834560156, + 0.0035893917083740234, + 0.0020330927800387144, + 0.002318799961358309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.003195063676685095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079810559749603 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079810559749603, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079810559749603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04563138633966446, + 0.841576099395752, + 0.011188565753400326, + 0.010499791242182255, + 0.009519848972558975, + 0.019771266728639603, + 0.013675482012331486, + 0.012173358350992203, + 0.022955289110541344, + 0.013008969835937023 + ], + "credit_map": [ + 0.002494585467502475, + 0.009995407424867153, + 0.0027523241005837917, + 0.0022410035599023104, + 0.0163832176476717, + 0.005019730422645807, + 0.002633752068504691, + 0.004111364483833313, + 0.03873884677886963, + 0.3465111553668976, + 0.007861566729843616, + 0.015071812085807323, + 0.35181817412376404, + 0.05959979072213173, + 0.01261660736054182, + 0.009358098730444908, + 0.0011391181033104658, + 0.0015576526056975126, + 0.0019387664506211877, + 0.0015350555768236518, + 0.0018012181390076876, + 0.000996521906927228, + 0.001494149793870747, + 0.0007260822458192706, + 0.0016187760047614574, + 0.0012816123198717833, + 0.001055428059771657, + 0.0016106965485960245, + 0.0016834704438224435, + 0.0012291967868804932, + 0.001304060686379671, + 0.0007165494607761502, + 0.0013774181716144085, + 0.002449183026328683, + 0.0008471512119285762, + 0.0010498786577954888, + 0.0010207169689238071, + 0.0012802110286429524, + 0.00080641865497455, + 0.0006888706702739, + 0.0028279414400458336, + 0.0038789338432252407, + 0.003360644681379199, + 0.001823127269744873, + 0.003415479324758053, + 0.001841115765273571, + 0.001506524858996272, + 0.0011174981482326984, + 0.0021754628978669643, + 0.0032564119901508093, + 0.002015934558585286, + 0.001037033973261714, + 0.0012567637022584677, + 0.001679257140494883, + 0.0009119197493419051, + 0.0013426963705569506, + 0.001818506047129631, + 0.0023554188665002584, + 0.0011825398541986942, + 0.0011677714064717293, + 0.0016625523567199707, + 0.0021331002935767174, + 0.0007866403320804238, + 0.0010668294271454215, + 0.003239405807107687, + 0.0025380514562129974, + 0.003282182151451707, + 0.00510661443695426, + 0.0019476704765111208, + 0.0027108159847557545, + 0.002058533951640129, + 0.002072012284770608, + 0.002088890876621008, + 0.001679159584455192, + 0.001786033040843904, + 0.001997870858758688, + 0.0012576315784826875, + 0.0018606528174132109, + 0.001168385031633079, + 0.0011703443014994264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.011079810559749603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a low angle with the camera tilting upwards at a slow, steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wilderness with withered dry grass and twisted leafless trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds or thousands of pitch-black crows visible roosting on the ground and dead branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crows suddenly flap their wings and surge upward into the sky forming a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky shrouded in thick swirling leaden clouds with a monochromatic hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197415672242641 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197415672242641, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197415672242641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10697577148675919, + 0.5488176941871643, + 0.13173605501651764, + 0.02453918196260929, + 0.013109014369547367, + 0.0145863126963377, + 0.02807687595486641, + 0.04799170792102814, + 0.04721179977059364, + 0.036955587565898895 + ], + "credit_map": [ + 0.0077208904549479485, + 0.0220539178699255, + 0.006602947134524584, + 0.006281338166445494, + 0.03724992275238037, + 0.011792371980845928, + 0.00885010790079832, + 0.006424261257052422, + 0.026094356551766396, + 0.2276022881269455, + 0.006178779527544975, + 0.010373713448643684, + 0.2263457328081131, + 0.03721826523542404, + 0.00789693370461464, + 0.007107575889676809, + 0.022283757105469704, + 0.02470623143017292, + 0.012488789856433868, + 0.007279821205884218, + 0.03136080130934715, + 0.018915731459856033, + 0.008572028949856758, + 0.006128885317593813, + 0.00440827663987875, + 0.0027051051147282124, + 0.003820689395070076, + 0.0024820533581078053, + 0.0032194883096963167, + 0.0033025206066668034, + 0.0028352283407002687, + 0.0017658197320997715, + 0.0015598656609654427, + 0.001898022834211588, + 0.001603142824023962, + 0.001746782218106091, + 0.0018439441919326782, + 0.0019590952433645725, + 0.001196518074721098, + 0.001301642507314682, + 0.0017863870598375797, + 0.0017691830871626735, + 0.0024470880161970854, + 0.0018143999623134732, + 0.0017909520538523793, + 0.0013545665424317122, + 0.001866291742771864, + 0.0017574417870491743, + 0.005045298486948013, + 0.008511603809893131, + 0.003146156668663025, + 0.0020776682067662477, + 0.0022821209859102964, + 0.0030030529014766216, + 0.0017806383548304439, + 0.0022303343284875154, + 0.008699582889676094, + 0.013259678147733212, + 0.005059563089162111, + 0.00441822549328208, + 0.004959359765052795, + 0.006038601975888014, + 0.00245665037073195, + 0.0031000408343970776, + 0.008232240565121174, + 0.013184734620153904, + 0.00804438628256321, + 0.004892617929726839, + 0.004558516200631857, + 0.0033197978045791388, + 0.002431014087051153, + 0.002548488322645426, + 0.008131506852805614, + 0.006974628195166588, + 0.007225738372653723, + 0.0037259573582559824, + 0.003742569126188755, + 0.0031848084181547165, + 0.001638180110603571, + 0.002332196570932865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.014197415672242641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_tensor([1154, 1154])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_tensor([1154, 1154])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad6e4f903932ecdb1d01330c974f7e9011bfd0f8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_tensor([1154, 1154])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-altitude, black-and-white drone view slowly inched toward the center of a vast volcanic crater, capturing its majestic and terrifying presence. The crater's jagged rims frame a deep, nearly black abyss, surrounded by a gray rock mass riddled with cracks and gullies. Suddenly, the center of the crater violently rips open, and a stream of searing white magma erupts explosively. Like a vast, luminous fountain, it shoots skyward, hurling countless molten spheres into the air. With a relentless, almost overexposed brightness, the viscous lava spills from the crater in streams of brilliant white, slowly but surely carving new paths across the vast, desolate, dark, ash-covered slopes. Simultaneously, a massive, billowing column of smoke and ash rises, its intricate gray textures creating a stark visual contrast with the fiery magma below. The entire scene is completely illuminated by the light of the magma itself, casting violently dancing shadows and creating a primal, apocalyptic atmosphere full of the irresistible and awe-inspiring power of nature.", + "seed": 11022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11022_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23027753350324928, + "vlm_logprob": -0.23027753350324928, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.23027753350324928, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9142042398452759 + ], + "opposite_logprob_tokens": [ + -0.5392042398452759 + ], + "logprob_reward": -0.9142042398452759, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9142042398452759, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.27501949667930603, + 0.08069860190153122, + 0.08615773171186447, + 0.07766680419445038, + 0.15417709946632385, + 0.09106458723545074, + 0.06933190673589706, + 0.10162628442049026, + 0.03462805971503258, + 0.029629463329911232 + ], + "credit_map": [ + 0.030115360394120216, + 0.0378468856215477, + 0.045835431665182114, + 0.03435308486223221, + 0.041034996509552, + 0.030742021277546883, + 0.02967015840113163, + 0.025421565398573875, + 0.0068112146109342575, + 0.01818203367292881, + 0.009895178489387035, + 0.00617713388055563, + 0.012552556581795216, + 0.0049745975993573666, + 0.011922614648938179, + 0.01018327847123146, + 0.009573006071150303, + 0.006117213983088732, + 0.013431564904749393, + 0.010867305099964142, + 0.010958357714116573, + 0.008317286148667336, + 0.014402845874428749, + 0.012490144930779934, + 0.007108921185135841, + 0.008475299924612045, + 0.008116756565868855, + 0.005891066510230303, + 0.006089914124459028, + 0.01208111084997654, + 0.01701471209526062, + 0.012889024801552296, + 0.013006151653826237, + 0.04064474254846573, + 0.01551301870495081, + 0.015671079978346825, + 0.01572892628610134, + 0.02710479497909546, + 0.01608921028673649, + 0.01041917223483324, + 0.0075155324302613735, + 0.00931398756802082, + 0.012222602032124996, + 0.01044012326747179, + 0.008750644512474537, + 0.012792237102985382, + 0.016472218558192253, + 0.013557239435613155, + 0.007706997916102409, + 0.008066540583968163, + 0.011780987493693829, + 0.011351018212735653, + 0.004915371537208557, + 0.00846350472420454, + 0.009320627897977829, + 0.0077268569730222225, + 0.012444334104657173, + 0.010981717146933079, + 0.013249595649540424, + 0.011086002923548222, + 0.008780386298894882, + 0.01533440686762333, + 0.017413459718227386, + 0.012336374260485172, + 0.003199945203959942, + 0.003425339236855507, + 0.004108343739062548, + 0.0031946366652846336, + 0.0031829127110540867, + 0.006368562579154968, + 0.006083474028855562, + 0.005064843688160181, + 0.002383075188845396, + 0.0033883852884173393, + 0.003571705659851432, + 0.00521489791572094, + 0.0030606621876358986, + 0.00601234519854188, + 0.003178416984155774, + 0.0028199758380651474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.9142042398452759 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20167528092861176 + ], + "opposite_logprob_tokens": [ + -1.701675295829773 + ], + "logprob_reward": -0.20167528092861176, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20167528092861176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25418752431869507, + 0.20239275693893433, + 0.054649945348501205, + 0.06452353298664093, + 0.062042757868766785, + 0.146550714969635, + 0.04261814430356026, + 0.04565520957112312, + 0.02237825281918049, + 0.10500115901231766 + ], + "credit_map": [ + 0.024676207453012466, + 0.03491285815834999, + 0.03005174547433853, + 0.026670385152101517, + 0.04833873361349106, + 0.03652071952819824, + 0.03191926330327988, + 0.021097611635923386, + 0.01920047402381897, + 0.0637689009308815, + 0.02474786341190338, + 0.010910775512456894, + 0.022023649886250496, + 0.009232325479388237, + 0.014319966547191143, + 0.038188811391592026, + 0.006632049567997456, + 0.004051268566399813, + 0.008771141991019249, + 0.0064595467410981655, + 0.00927315466105938, + 0.004764987155795097, + 0.008767513558268547, + 0.005930284503847361, + 0.006878824904561043, + 0.006273849867284298, + 0.005843857768923044, + 0.005298088304698467, + 0.006875221151858568, + 0.013561086729168892, + 0.013400055468082428, + 0.006392550654709339, + 0.006241778377443552, + 0.006563643924891949, + 0.010211171582341194, + 0.005145895294845104, + 0.0073521980084478855, + 0.00783362053334713, + 0.009218632243573666, + 0.00947581510990858, + 0.012889903038740158, + 0.011815846897661686, + 0.02431405335664749, + 0.012084420770406723, + 0.013659356161952019, + 0.01987338252365589, + 0.02407067082822323, + 0.027843080461025238, + 0.0037068945821374655, + 0.005537747871130705, + 0.004913186188787222, + 0.006061688996851444, + 0.004420110955834389, + 0.0047629051841795444, + 0.008258841000497341, + 0.00495676975697279, + 0.004566499963402748, + 0.00445434357970953, + 0.005035444162786007, + 0.006402868311852217, + 0.003686564974486828, + 0.006956512574106455, + 0.007522667292505503, + 0.0070303063839674, + 0.0027133412659168243, + 0.0021232166327536106, + 0.002176924142986536, + 0.0025573745369911194, + 0.002330784685909748, + 0.0046158055774867535, + 0.002981941681355238, + 0.0028788652271032333, + 0.009538053534924984, + 0.014199997298419476, + 0.00920112244784832, + 0.023145094513893127, + 0.010215749964118004, + 0.016656257212162018, + 0.011759253218770027, + 0.010285628028213978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.20167528092861176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815908521413803 + ], + "opposite_logprob_tokens": [ + -3.529815912246704 + ], + "logprob_reward": -0.029815908521413803, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815908521413803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07772066444158554, + 0.036218006163835526, + 0.05761657655239105, + 0.2211574763059616, + 0.1543177217245102, + 0.0759999006986618, + 0.13297420740127563, + 0.17635034024715424, + 0.043211523443460464, + 0.024433558806777 + ], + "credit_map": [ + 0.008132156915962696, + 0.011387010104954243, + 0.011432981118559837, + 0.01040615700185299, + 0.013659784570336342, + 0.008119655773043633, + 0.008490489795804024, + 0.006092427764087915, + 0.0031736581586301327, + 0.0051954458467662334, + 0.004903844557702541, + 0.004254699684679508, + 0.006493986584246159, + 0.004205022472888231, + 0.004699347540736198, + 0.003292004344984889, + 0.006663967855274677, + 0.0038204321172088385, + 0.00843728706240654, + 0.007307685445994139, + 0.0070260572247207165, + 0.004569808021187782, + 0.009580262936651707, + 0.01021107379347086, + 0.019799552857875824, + 0.022213827818632126, + 0.025345578789711, + 0.017442280426621437, + 0.017048005014657974, + 0.041676025837659836, + 0.0449097715318203, + 0.032722439616918564, + 0.013093133457005024, + 0.03729867562651634, + 0.0157476793974638, + 0.015697907656431198, + 0.019042279571294785, + 0.027573423460125923, + 0.014223592355847359, + 0.011641029268503189, + 0.009408529847860336, + 0.007133618928492069, + 0.008958582766354084, + 0.008064297027885914, + 0.007831920869648457, + 0.008726620115339756, + 0.014128414914011955, + 0.011747916229069233, + 0.011073929257690907, + 0.015719620510935783, + 0.015661679208278656, + 0.01764840818941593, + 0.014554839581251144, + 0.01550255436450243, + 0.025726307183504105, + 0.01708686724305153, + 0.017726892605423927, + 0.024418732151389122, + 0.023503463715314865, + 0.031757060438394547, + 0.013961621560156345, + 0.015596650540828705, + 0.03431781753897667, + 0.015068097971379757, + 0.006089681293815374, + 0.004708126652985811, + 0.004684052430093288, + 0.005614321678876877, + 0.004207897465676069, + 0.00574225839227438, + 0.00517854792997241, + 0.0069866362027823925, + 0.004374214448034763, + 0.0026827433612197638, + 0.002169274492189288, + 0.002990039763972163, + 0.0025236867368221283, + 0.0029530134052038193, + 0.003780520986765623, + 0.002960066543892026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.029815908521413803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024964616168290377 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.0024964616168290377, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024964616168290377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08131740987300873, + 0.23396527767181396, + 0.17689551413059235, + 0.2157745212316513, + 0.06942480057477951, + 0.06757886707782745, + 0.036758989095687866, + 0.05220937356352806, + 0.03982393816113472, + 0.026251336559653282 + ], + "credit_map": [ + 0.0069714258424937725, + 0.009063657373189926, + 0.013696236535906792, + 0.011848773807287216, + 0.011003399267792702, + 0.011211942881345749, + 0.008518795482814312, + 0.009003188461065292, + 0.02520330622792244, + 0.08157061040401459, + 0.027774209156632423, + 0.01712178997695446, + 0.025451919063925743, + 0.007207466289401054, + 0.016373716294765472, + 0.03326224535703659, + 0.019358830526471138, + 0.012980100698769093, + 0.02996211312711239, + 0.021949606016278267, + 0.02262364700436592, + 0.015504352748394012, + 0.025761425495147705, + 0.028755415230989456, + 0.018383944407105446, + 0.022644704207777977, + 0.020437663421034813, + 0.015435545705258846, + 0.016997650265693665, + 0.029299752786755562, + 0.05208670720458031, + 0.04048854485154152, + 0.005273942369967699, + 0.013018068857491016, + 0.006628900300711393, + 0.007494054734706879, + 0.009551145136356354, + 0.009468816220760345, + 0.011510676704347134, + 0.00647919112816453, + 0.006101991515606642, + 0.00644331518560648, + 0.006719814147800207, + 0.006547784898430109, + 0.006593883037567139, + 0.008531946688890457, + 0.01549664419144392, + 0.011143490672111511, + 0.004077920690178871, + 0.004096054472029209, + 0.004307781346142292, + 0.00586320785805583, + 0.004354030825197697, + 0.004017923027276993, + 0.005684243515133858, + 0.004357829224318266, + 0.005243253894150257, + 0.006023063324391842, + 0.0057706027291715145, + 0.006612055003643036, + 0.005231591407209635, + 0.0063508958555758, + 0.011425703763961792, + 0.005552212242037058, + 0.004161523655056953, + 0.00407381821423769, + 0.00617096945643425, + 0.0059853121638298035, + 0.0038698771968483925, + 0.005003938917070627, + 0.005913355853408575, + 0.004645143169909716, + 0.0028234333731234074, + 0.0025594329927116632, + 0.003585502039641142, + 0.003550187451764941, + 0.0029734789859503508, + 0.004128818400204182, + 0.0031881798058748245, + 0.003442303976044059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.0024964616168290377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1428312212228775, + 0.2712862193584442, + 0.14444391429424286, + 0.07211010903120041, + 0.07295309752225876, + 0.04673171788454056, + 0.03546200692653656, + 0.03858568146824837, + 0.0876249298453331, + 0.08797112852334976 + ], + "credit_map": [ + 0.013272464275360107, + 0.015889475122094154, + 0.021488294005393982, + 0.01769001968204975, + 0.019732993096113205, + 0.018679285421967506, + 0.0198344886302948, + 0.016244199126958847, + 0.025437256321310997, + 0.07852520793676376, + 0.03250930830836296, + 0.021163152530789375, + 0.030885666608810425, + 0.013661893084645271, + 0.02030058391392231, + 0.04880314692854881, + 0.01610376499593258, + 0.010593252256512642, + 0.023038635030388832, + 0.021259156987071037, + 0.021409200504422188, + 0.011096673086285591, + 0.026651758700609207, + 0.014291476458311081, + 0.0075025614351034164, + 0.0053783562034368515, + 0.009288865141570568, + 0.004361581057310104, + 0.006411346141248941, + 0.025000296533107758, + 0.008271987549960613, + 0.005895114038139582, + 0.006642162799835205, + 0.01602606102824211, + 0.00914714764803648, + 0.006601554341614246, + 0.005330332089215517, + 0.015531238168478012, + 0.007586634252220392, + 0.006087964866310358, + 0.004567825701087713, + 0.004095544572919607, + 0.007545253727585077, + 0.005154691636562347, + 0.0037143612280488014, + 0.005368231330066919, + 0.0075194453820586205, + 0.008766364306211472, + 0.003083029994741082, + 0.004363886080682278, + 0.004440301097929478, + 0.006500687450170517, + 0.0034271811600774527, + 0.00415818439796567, + 0.005683779250830412, + 0.0038049572613090277, + 0.004183562006801367, + 0.004132366273552179, + 0.004259698558598757, + 0.005371845792979002, + 0.003308997955173254, + 0.0043145352974534035, + 0.0075099351815879345, + 0.00550474226474762, + 0.007048959843814373, + 0.01189990434795618, + 0.0077368407510221004, + 0.020459342747926712, + 0.008320963941514492, + 0.008412722498178482, + 0.015428039245307446, + 0.008318157866597176, + 0.014347623102366924, + 0.010043424554169178, + 0.0102079464122653, + 0.016451293602585793, + 0.007956258952617645, + 0.011881370097398758, + 0.007675223983824253, + 0.009407984092831612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.003195776604115963 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude, black-and-white drone view slowly inched toward the center of a vast volcanic crater, capturing its majestic and terrifying presence. The crater's jagged rims frame a deep, nearly black abyss, surrounded by a gray rock mass riddled with cracks and gullies. Suddenly, the center of the crater violently rips open, and a stream of searing white magma erupts explosively. Like a vast, luminous fountain, it shoots skyward, hurling countless molten spheres into the air. With a relentless, almost overexposed brightness, the viscous lava spills from the crater in streams of brilliant white, slowly but surely carving new paths across the vast, desolate, dark, ash-covered slopes. Simultaneously, a massive, billowing column of smoke and ash rises, its intricate gray textures creating a stark visual contrast with the fiery magma below. The entire scene is completely illuminated by the light of the magma itself, casting violently dancing shadows and creating a primal, apocalyptic atmosphere full of the irresistible and awe-inspiring power of nature.", + "seed": 11023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11023_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0339745931327343, + "vlm_logprob": -0.0339745931327343, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0339745931327343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11439578980207443 + ], + "opposite_logprob_tokens": [ + -2.4893958568573 + ], + "logprob_reward": -0.11439578980207443, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11439578980207443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0791119858622551, + 0.0683426782488823, + 0.08362814038991928, + 0.04474500194191933, + 0.15528768301010132, + 0.22722560167312622, + 0.044052399694919586, + 0.05906841531395912, + 0.039573539048433304, + 0.19896458089351654 + ], + "credit_map": [ + 0.006548888515681028, + 0.008607929572463036, + 0.011082129552960396, + 0.00786704383790493, + 0.013861271552741528, + 0.009994170628488064, + 0.009776157326996326, + 0.011374395340681076, + 0.007415552623569965, + 0.006982201710343361, + 0.009819726459681988, + 0.006040836218744516, + 0.014087938703596592, + 0.008971861563622952, + 0.007472244091331959, + 0.007552315015345812, + 0.00463796965777874, + 0.004141069017350674, + 0.005579391960054636, + 0.003637437243014574, + 0.029917435720562935, + 0.008361758664250374, + 0.005646624136716127, + 0.02170645445585251, + 0.0060880440287292, + 0.005316464230418205, + 0.0042637307196855545, + 0.003792861709371209, + 0.0062584164552390575, + 0.008723455481231213, + 0.004099546931684017, + 0.006202484481036663, + 0.012912687845528126, + 0.05541609227657318, + 0.008459393866360188, + 0.005692935083061457, + 0.044232867658138275, + 0.01322384737432003, + 0.0060644978657364845, + 0.009285354055464268, + 0.02794363722205162, + 0.019260348752141, + 0.035928670316934586, + 0.029807059094309807, + 0.027511630207300186, + 0.03550158068537712, + 0.02620932087302208, + 0.025063367560505867, + 0.005245677661150694, + 0.004297775216400623, + 0.005718240514397621, + 0.004026955459266901, + 0.006997907999902964, + 0.007210763171315193, + 0.0054818023927509785, + 0.005073277745395899, + 0.0059060039930045605, + 0.0058959536254405975, + 0.005682158283889294, + 0.005209807772189379, + 0.009569576010107994, + 0.012266967445611954, + 0.006318594794720411, + 0.008219351060688496, + 0.004089830443263054, + 0.003670099889859557, + 0.0035470251459628344, + 0.004407580941915512, + 0.0061812265776097775, + 0.006464283913373947, + 0.005587527994066477, + 0.0056259650737047195, + 0.03090890496969223, + 0.012930275872349739, + 0.020980484783649445, + 0.008849846199154854, + 0.018759647384285927, + 0.05568118765950203, + 0.01004636287689209, + 0.04080788418650627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.11439578980207443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006006879266351461 + ], + "opposite_logprob_tokens": [ + -5.131006717681885 + ], + "logprob_reward": -0.006006879266351461, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006006879266351461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33858609199523926, + 0.045173585414886475, + 0.039009567350149155, + 0.17763832211494446, + 0.07112564891576767, + 0.048544544726610184, + 0.0404045507311821, + 0.08233758062124252, + 0.045806825160980225, + 0.11137328296899796 + ], + "credit_map": [ + 0.02959335781633854, + 0.04064464941620827, + 0.05568163841962814, + 0.04710440710186958, + 0.046024855226278305, + 0.04325271397829056, + 0.03556326404213905, + 0.040721211582422256, + 0.003830048255622387, + 0.004782155156135559, + 0.0077621182426810265, + 0.005603146739304066, + 0.0051345303654670715, + 0.006803024094551802, + 0.00517518213018775, + 0.006083382293581963, + 0.0028996437322348356, + 0.0037336309906095266, + 0.0034928584937006235, + 0.003640158800408244, + 0.007910958491265774, + 0.009244323708117008, + 0.004405794199556112, + 0.003682199399918318, + 0.018809646368026733, + 0.02894553542137146, + 0.018025942146778107, + 0.025266224518418312, + 0.024812087416648865, + 0.027051113545894623, + 0.018299255520105362, + 0.01642853207886219, + 0.007308862172067165, + 0.014392460696399212, + 0.008101912215352058, + 0.003488363465294242, + 0.01150320190936327, + 0.013299728743731976, + 0.0062926337122917175, + 0.006738487631082535, + 0.005873383488506079, + 0.004195376764982939, + 0.005549450404942036, + 0.0062322597950696945, + 0.009679371491074562, + 0.006039090920239687, + 0.006648917216807604, + 0.004326697438955307, + 0.0039076958782970905, + 0.004921053536236286, + 0.004658970050513744, + 0.004842715337872505, + 0.007210100069642067, + 0.005645426455885172, + 0.004415914881974459, + 0.0048026759177446365, + 0.010010899044573307, + 0.009192812256515026, + 0.009765286929905415, + 0.006527220830321312, + 0.017168594524264336, + 0.009824680164456367, + 0.008134634234011173, + 0.011713454499840736, + 0.005217911209911108, + 0.004868977703154087, + 0.0040105367079377174, + 0.002859831554815173, + 0.011367416940629482, + 0.008566849865019321, + 0.004041121806949377, + 0.004874181002378464, + 0.015808533877134323, + 0.005655006505548954, + 0.03108952008187771, + 0.006292961537837982, + 0.0053323982283473015, + 0.028690766543149948, + 0.005899819079786539, + 0.012604287825524807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006006879266351461 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307185113430023 + ], + "opposite_logprob_tokens": [ + -3.168071746826172 + ], + "logprob_reward": -0.04307185113430023, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307185113430023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10569693893194199, + 0.113706074655056, + 0.13586199283599854, + 0.07169067859649658, + 0.1462513655424118, + 0.11525178700685501, + 0.05637387931346893, + 0.04628056660294533, + 0.025483388453722, + 0.1834034025669098 + ], + "credit_map": [ + 0.011128024198114872, + 0.012345780618488789, + 0.018665794283151627, + 0.015066509135067463, + 0.012871929444372654, + 0.011225991882383823, + 0.011505820788443089, + 0.012887096963822842, + 0.010853867046535015, + 0.01445469819009304, + 0.015432651154696941, + 0.008979753591120243, + 0.02392999269068241, + 0.01133826095610857, + 0.011995856650173664, + 0.01672099530696869, + 0.008955900557339191, + 0.01478161383420229, + 0.011726860888302326, + 0.00834873877465725, + 0.020269451662898064, + 0.02791411802172661, + 0.010296110063791275, + 0.033569201827049255, + 0.006896118633449078, + 0.009795460850000381, + 0.00734670041128993, + 0.010760357603430748, + 0.008636603131890297, + 0.01288228202611208, + 0.007682660594582558, + 0.007690494414418936, + 0.015578518621623516, + 0.02911302074790001, + 0.013848554342985153, + 0.009590393863618374, + 0.022460417822003365, + 0.02657128870487213, + 0.014328813180327415, + 0.014760353602468967, + 0.014159748330712318, + 0.010083187371492386, + 0.011099903844296932, + 0.015879346057772636, + 0.02316279150545597, + 0.012394924648106098, + 0.01767105422914028, + 0.010800831019878387, + 0.008306102827191353, + 0.004794536158442497, + 0.007633634842932224, + 0.007291393354535103, + 0.007092283573001623, + 0.007613821420818567, + 0.006323784589767456, + 0.007318318821489811, + 0.005147927440702915, + 0.004666951950639486, + 0.006715425290167332, + 0.003160452703014016, + 0.009748064912855625, + 0.007787609472870827, + 0.00352192809805274, + 0.00553220696747303, + 0.0017568658804520965, + 0.0024899267591536045, + 0.0017118226969614625, + 0.002238551154732704, + 0.005584982223808765, + 0.004631249699741602, + 0.0038041803054511547, + 0.0032658090349286795, + 0.04714249446988106, + 0.01095028966665268, + 0.020432936027646065, + 0.008158927783370018, + 0.012671071104705334, + 0.05420946329832077, + 0.007959206588566303, + 0.021879013627767563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.04307185113430023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032007675617933273 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032007675617933273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032007675617933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06706129014492035, + 0.050157688558101654, + 0.07438521087169647, + 0.32670894265174866, + 0.26746806502342224, + 0.09283234924077988, + 0.04068522900342941, + 0.018466079607605934, + 0.03111027367413044, + 0.031124824658036232 + ], + "credit_map": [ + 0.0060688224621117115, + 0.0065194396302104, + 0.009627943858504295, + 0.008508080616593361, + 0.011067510582506657, + 0.008471643552184105, + 0.008424405939877033, + 0.008373448625206947, + 0.0048822071403265, + 0.005408833734691143, + 0.006057507824152708, + 0.0048767924308776855, + 0.008900856599211693, + 0.006221431773155928, + 0.007307087536901236, + 0.006502973381429911, + 0.004442057106643915, + 0.005435312166810036, + 0.004535997286438942, + 0.0039791539311409, + 0.01337832398712635, + 0.01520039327442646, + 0.006301999092102051, + 0.02111198566854, + 0.03342009335756302, + 0.05903907120227814, + 0.030374327674508095, + 0.04509414732456207, + 0.04573056474328041, + 0.048318956047296524, + 0.033354684710502625, + 0.03137713298201561, + 0.02074310928583145, + 0.08245697617530823, + 0.018906261771917343, + 0.011017563752830029, + 0.06855197995901108, + 0.03371013328433037, + 0.012720274738967419, + 0.01936177723109722, + 0.00990324653685093, + 0.008576846681535244, + 0.013987093232572079, + 0.013030324131250381, + 0.016826491802930832, + 0.010833317413926125, + 0.010092014446854591, + 0.009583016857504845, + 0.004103906452655792, + 0.004402998834848404, + 0.004642100539058447, + 0.005381536204367876, + 0.007036146707832813, + 0.005290770437568426, + 0.003946382086724043, + 0.005881392862647772, + 0.002106526866555214, + 0.0020752691198140383, + 0.0017060275422409177, + 0.0018361465772613883, + 0.004177778493613005, + 0.0026018412318080664, + 0.001832192181609571, + 0.0021303000394254923, + 0.0029480974189937115, + 0.0037571494467556477, + 0.0027275290340185165, + 0.004055935889482498, + 0.004862785339355469, + 0.0038149196188896894, + 0.005584066733717918, + 0.003359792288392782, + 0.00528391869738698, + 0.0016695103840902448, + 0.002753807231783867, + 0.0012172376736998558, + 0.0025870015379041433, + 0.010420705191791058, + 0.0012282063253223896, + 0.005964439827948809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.0032007675617933273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031976778991520405 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031976778991520405, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031976778991520405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12803414463996887, + 0.04946319758892059, + 0.06058016046881676, + 0.258077472448349, + 0.17047874629497528, + 0.07180541753768921, + 0.049170129001140594, + 0.0716349333524704, + 0.04431615024805069, + 0.09643964469432831 + ], + "credit_map": [ + 0.010475372895598412, + 0.017490938305854797, + 0.01999903842806816, + 0.01685907319188118, + 0.018603330478072166, + 0.016833702102303505, + 0.013261674903333187, + 0.01451100967824459, + 0.005887366831302643, + 0.005765664856880903, + 0.006676462013274431, + 0.004401069134473801, + 0.007504642475396395, + 0.0067279222421348095, + 0.005721723195165396, + 0.006778346374630928, + 0.0046950578689575195, + 0.006938953883945942, + 0.005191221833229065, + 0.004906385205686092, + 0.010904605500400066, + 0.008494004607200623, + 0.007112985476851463, + 0.012336947955191135, + 0.027433620765805244, + 0.0464172437787056, + 0.025046637281775475, + 0.035952772945165634, + 0.0360824353992939, + 0.03718746826052666, + 0.026330498978495598, + 0.023626796901226044, + 0.01883610151708126, + 0.04721810296177864, + 0.015378842130303383, + 0.007290284615010023, + 0.032125502824783325, + 0.02675139345228672, + 0.007839200086891651, + 0.015039310790598392, + 0.008940249681472778, + 0.00606690114364028, + 0.007199454121291637, + 0.009975843131542206, + 0.01396070048213005, + 0.009523064829409122, + 0.008126463741064072, + 0.008012739941477776, + 0.005835593678057194, + 0.00501724798232317, + 0.006565597839653492, + 0.006220017094165087, + 0.008261383511126041, + 0.006546251941472292, + 0.006288088392466307, + 0.0044359443709254265, + 0.006916423793882132, + 0.008223806507885456, + 0.005276419222354889, + 0.009080834686756134, + 0.01750345714390278, + 0.00993520487099886, + 0.005917278118431568, + 0.008781509473919868, + 0.004437892697751522, + 0.006440947763621807, + 0.0044438657350838184, + 0.004327756818383932, + 0.008789901621639729, + 0.006347344722598791, + 0.004909969866275787, + 0.004618472885340452, + 0.0139867402613163, + 0.0035393836442381144, + 0.03327620029449463, + 0.0050476728938519955, + 0.005093688610941172, + 0.01380101591348648, + 0.0034905853681266308, + 0.018204359337687492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0031976778991520405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude, black-and-white drone view slowly inched toward the center of a vast volcanic crater, capturing its majestic and terrifying presence. The crater's jagged rims frame a deep, nearly black abyss, surrounded by a gray rock mass riddled with cracks and gullies. Suddenly, the center of the crater violently rips open, and a stream of searing white magma erupts explosively. Like a vast, luminous fountain, it shoots skyward, hurling countless molten spheres into the air. With a relentless, almost overexposed brightness, the viscous lava spills from the crater in streams of brilliant white, slowly but surely carving new paths across the vast, desolate, dark, ash-covered slopes. Simultaneously, a massive, billowing column of smoke and ash rises, its intricate gray textures creating a stark visual contrast with the fiery magma below. The entire scene is completely illuminated by the light of the magma itself, casting violently dancing shadows and creating a primal, apocalyptic atmosphere full of the irresistible and awe-inspiring power of nature.", + "seed": 11024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11024_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06688958858139812, + "vlm_logprob": -0.06688958858139812, + "vlm_yes_no_margin": 3.949999976158142, + "vlm_token_credit": -0.06688958858139812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15182049572467804 + ], + "opposite_logprob_tokens": [ + -2.151820421218872 + ], + "logprob_reward": -0.15182049572467804, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.15182049572467804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25955861806869507, + 0.0670728087425232, + 0.15456578135490417, + 0.09201648086309433, + 0.08181886374950409, + 0.056638315320014954, + 0.0545426681637764, + 0.08682595193386078, + 0.08243190497159958, + 0.06452861428260803 + ], + "credit_map": [ + 0.02674471214413643, + 0.035220496356487274, + 0.03759796544909477, + 0.036695171147584915, + 0.036075856536626816, + 0.03758561983704567, + 0.026160085573792458, + 0.02347872219979763, + 0.008330953307449818, + 0.006601743865758181, + 0.010799350216984749, + 0.010020545683801174, + 0.009231497533619404, + 0.00820235162973404, + 0.00522581534460187, + 0.008660541847348213, + 0.013225892558693886, + 0.016320854425430298, + 0.027611732482910156, + 0.026447342708706856, + 0.020195171236991882, + 0.01959306001663208, + 0.013077057898044586, + 0.018094658851623535, + 0.011852459982037544, + 0.009660936892032623, + 0.01493862271308899, + 0.01229079719632864, + 0.010050354525446892, + 0.013984329998493195, + 0.007030745502561331, + 0.012208235450088978, + 0.010602215304970741, + 0.009759224951267242, + 0.010249326005578041, + 0.010721776634454727, + 0.0150584252551198, + 0.010630164295434952, + 0.00877782329916954, + 0.006019904278218746, + 0.00575815187767148, + 0.00605896906927228, + 0.006341616157442331, + 0.00932804774492979, + 0.006408436223864555, + 0.0066183824092149734, + 0.007605025079101324, + 0.008519689552485943, + 0.006111182738095522, + 0.005249876994639635, + 0.009283010847866535, + 0.004257377702742815, + 0.006257005501538515, + 0.010754263959825039, + 0.007225124165415764, + 0.005404822528362274, + 0.008783962577581406, + 0.011708196252584457, + 0.011758013628423214, + 0.006998321507126093, + 0.01355583406984806, + 0.018429309129714966, + 0.006883990950882435, + 0.008708326146006584, + 0.011307156644761562, + 0.016498079523444176, + 0.007688801735639572, + 0.013916506431996822, + 0.009819931350648403, + 0.008167929016053677, + 0.009043769910931587, + 0.005989728029817343, + 0.0082828588783741, + 0.011593088507652283, + 0.010862615890800953, + 0.008320062421262264, + 0.00684376060962677, + 0.007092752028256655, + 0.006150882225483656, + 0.005382599774748087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.15182049572467804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038763284683228 + ], + "opposite_logprob_tokens": [ + -2.3503875732421875 + ], + "logprob_reward": -0.10038763284683228, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038763284683228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11973539739847183, + 0.0823553279042244, + 0.08057492226362228, + 0.10032914578914642, + 0.25299715995788574, + 0.06738739460706711, + 0.022879472002387047, + 0.028425397351384163, + 0.1687777042388916, + 0.07653799653053284 + ], + "credit_map": [ + 0.011308062821626663, + 0.021863555535674095, + 0.0148671455681324, + 0.017352178692817688, + 0.013138692826032639, + 0.015456130728125572, + 0.012686286121606827, + 0.0130633395165205, + 0.010626518167555332, + 0.008220266550779343, + 0.013028656132519245, + 0.009413317777216434, + 0.010242341086268425, + 0.011853812262415886, + 0.007734918966889381, + 0.011235498823225498, + 0.008664486929774284, + 0.008037817664444447, + 0.014310985803604126, + 0.010937039740383625, + 0.010982404462993145, + 0.009205607697367668, + 0.006090471986681223, + 0.012346111238002777, + 0.011442135088145733, + 0.010453520342707634, + 0.015214243903756142, + 0.014294762164354324, + 0.010148403234779835, + 0.012336613610386848, + 0.012450836598873138, + 0.013988630846142769, + 0.026921112090349197, + 0.0302593931555748, + 0.03324456512928009, + 0.03621339425444603, + 0.04604839161038399, + 0.03938508406281471, + 0.024982120841741562, + 0.015943113714456558, + 0.007939208298921585, + 0.007551967166364193, + 0.007845336571335793, + 0.009495621547102928, + 0.010523774661123753, + 0.008856740780174732, + 0.007250294554978609, + 0.007924450561404228, + 0.002631550654768944, + 0.0022375669796019793, + 0.004781017545610666, + 0.003130884375423193, + 0.0028652562759816647, + 0.0027098250575363636, + 0.0026908128056675196, + 0.0018325583077967167, + 0.0030713530723005533, + 0.004539994522929192, + 0.004891389049589634, + 0.0032011785078793764, + 0.004178873263299465, + 0.0032133315689861774, + 0.0028135478496551514, + 0.0025157311465591192, + 0.02325138822197914, + 0.026464227586984634, + 0.02963418699800968, + 0.01982058584690094, + 0.020741986110806465, + 0.0200605858117342, + 0.018134742975234985, + 0.010670002549886703, + 0.008577316999435425, + 0.023249931633472443, + 0.00805298425257206, + 0.00907129980623722, + 0.011201310902833939, + 0.0062935263849794865, + 0.005879826378077269, + 0.004211804363876581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.10038763284683228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898933440446854 + ], + "opposite_logprob_tokens": [ + -2.578989267349243 + ], + "logprob_reward": -0.07898933440446854, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898933440446854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12480201572179794, + 0.11748633533716202, + 0.12068375200033188, + 0.06560472398996353, + 0.07867810875177383, + 0.06504819542169571, + 0.10670741647481918, + 0.0822688564658165, + 0.06701025366783142, + 0.17171035706996918 + ], + "credit_map": [ + 0.015032639726996422, + 0.017178788781166077, + 0.02440459094941616, + 0.014568159356713295, + 0.01331513375043869, + 0.015948709100484848, + 0.011388183571398258, + 0.012965802103281021, + 0.011841336265206337, + 0.011416413821280003, + 0.01689046062529087, + 0.02314029075205326, + 0.018620429560542107, + 0.012361493892967701, + 0.009519156068563461, + 0.013696737587451935, + 0.01081855222582817, + 0.01674877479672432, + 0.017265256494283676, + 0.019635498523712158, + 0.016381820663809776, + 0.01266658678650856, + 0.008703329600393772, + 0.018463920801877975, + 0.007682460360229015, + 0.007388857193291187, + 0.010838551446795464, + 0.010637986473739147, + 0.006673729047179222, + 0.008346626535058022, + 0.006302472669631243, + 0.007734033744782209, + 0.009690571576356888, + 0.009219993837177753, + 0.010694277472794056, + 0.00977875106036663, + 0.013067605905234814, + 0.011314640752971172, + 0.008393241092562675, + 0.00651901587843895, + 0.008345093578100204, + 0.007133471313863993, + 0.011002629064023495, + 0.010285439901053905, + 0.006399947218596935, + 0.008618618361651897, + 0.006799547467380762, + 0.006463441997766495, + 0.012790173292160034, + 0.012697895057499409, + 0.01461924146860838, + 0.008723819628357887, + 0.012272510677576065, + 0.019427336752414703, + 0.01633232645690441, + 0.009844113141298294, + 0.009658396244049072, + 0.009007872082293034, + 0.014928983524441719, + 0.010082527995109558, + 0.012838798575103283, + 0.012311110273003578, + 0.007204381749033928, + 0.006236779969185591, + 0.009048432111740112, + 0.009736239910125732, + 0.012737949378788471, + 0.008851439692080021, + 0.006414923816919327, + 0.007189041934907436, + 0.008805651217699051, + 0.0042265718802809715, + 0.019182967022061348, + 0.046892255544662476, + 0.020924905315041542, + 0.017553051933646202, + 0.022892899811267853, + 0.01935194432735443, + 0.01085642445832491, + 0.014055894687771797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07898933440446854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022054414730519056 + ], + "opposite_logprob_tokens": [ + -6.1272053718566895 + ], + "logprob_reward": -0.0022054414730519056, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022054414730519056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12137147784233093, + 0.1534331738948822, + 0.06915969401597977, + 0.07613182067871094, + 0.13042016327381134, + 0.05314847081899643, + 0.028823884204030037, + 0.06896994262933731, + 0.11611412465572357, + 0.18242722749710083 + ], + "credit_map": [ + 0.020352043211460114, + 0.022568484768271446, + 0.024912655353546143, + 0.012654118239879608, + 0.013982940465211868, + 0.009356100112199783, + 0.00821463018655777, + 0.0093305055052042, + 0.014958606101572514, + 0.015760527923703194, + 0.01986376941204071, + 0.03155357018113136, + 0.02414219081401825, + 0.015960058197379112, + 0.012667505070567131, + 0.01852695643901825, + 0.00798344612121582, + 0.0070934719406068325, + 0.009560145437717438, + 0.009589879773557186, + 0.007899888791143894, + 0.007433454040437937, + 0.005184791050851345, + 0.014414619654417038, + 0.007515766192227602, + 0.009414552710950375, + 0.009725417010486126, + 0.012941411696374416, + 0.009569965302944183, + 0.010177035816013813, + 0.007912739180028439, + 0.008874929510056973, + 0.0161732230335474, + 0.015308626927435398, + 0.01805785670876503, + 0.01766175962984562, + 0.021954810246825218, + 0.020138662308454514, + 0.012256763875484467, + 0.008868462406098843, + 0.007002739235758781, + 0.005055042915046215, + 0.007541253697127104, + 0.007396697532385588, + 0.006802388001233339, + 0.007008267566561699, + 0.005208907183259726, + 0.007133173290640116, + 0.0029868572019040585, + 0.0027879062108695507, + 0.0040300446562469006, + 0.0034735959488898516, + 0.004092792049050331, + 0.00485234847292304, + 0.004015927202999592, + 0.002584413392469287, + 0.008907699957489967, + 0.007740532048046589, + 0.015307403169572353, + 0.008339748717844486, + 0.007145344279706478, + 0.009202086366713047, + 0.006197422277182341, + 0.0061297039501369, + 0.016462447121739388, + 0.016618123278021812, + 0.019329968839883804, + 0.014231107197701931, + 0.016264166682958603, + 0.013471400365233421, + 0.012308330275118351, + 0.00742858462035656, + 0.0228805560618639, + 0.024239135906100273, + 0.0252072811126709, + 0.03178148344159126, + 0.02198145166039467, + 0.017834192141890526, + 0.020709289237856865, + 0.01779385283589363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.0022054414730519056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010450384579598904 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010450384579598904, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010450384579598904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18091249465942383, + 0.06406982243061066, + 0.05078115686774254, + 0.1736656278371811, + 0.1977161467075348, + 0.039293449372053146, + 0.02221613936126232, + 0.04315716773271561, + 0.08116274327039719, + 0.14702525734901428 + ], + "credit_map": [ + 0.018311383202672005, + 0.029927996918559074, + 0.03148625046014786, + 0.024943646043539047, + 0.021060116589069366, + 0.02063690684735775, + 0.016695376485586166, + 0.01785082370042801, + 0.005205192137509584, + 0.0068678888492286205, + 0.0082467095926404, + 0.011848986148834229, + 0.010530236177146435, + 0.008180591277778149, + 0.006160023156553507, + 0.007030191831290722, + 0.004509958438575268, + 0.00625996757298708, + 0.00737967761233449, + 0.007786404807120562, + 0.008375827223062515, + 0.006435430608689785, + 0.0043417769484221935, + 0.005692112259566784, + 0.01781085878610611, + 0.021689046174287796, + 0.024926649406552315, + 0.029012825340032578, + 0.02313585765659809, + 0.025436054915189743, + 0.012889382429420948, + 0.01876494474709034, + 0.01673453114926815, + 0.025896701961755753, + 0.020639823749661446, + 0.029619066044688225, + 0.04079195111989975, + 0.029097609221935272, + 0.01734245754778385, + 0.017594026401638985, + 0.00423714704811573, + 0.004079020116478205, + 0.004763220902532339, + 0.004428563639521599, + 0.005906383041292429, + 0.0065511055290699005, + 0.004062838852405548, + 0.00526517303660512, + 0.0017802190268412232, + 0.0025175928603857756, + 0.0030063448939472437, + 0.004228669218719006, + 0.0032194512896239758, + 0.002840474946424365, + 0.002532489597797394, + 0.0020908962469547987, + 0.004225685261189938, + 0.004750834312289953, + 0.008015019819140434, + 0.004325852729380131, + 0.005075329449027777, + 0.008762119337916374, + 0.004044418688863516, + 0.003957911394536495, + 0.01263366173952818, + 0.01050059124827385, + 0.0099134286865592, + 0.014068014919757843, + 0.008722036145627499, + 0.009958194568753242, + 0.009654674679040909, + 0.005712134297937155, + 0.01485899556428194, + 0.03941328078508377, + 0.011361418291926384, + 0.020913954824209213, + 0.02016058750450611, + 0.010523010976612568, + 0.020061613991856575, + 0.009732402861118317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0010450384579598904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude, black-and-white drone view slowly inched toward the center of a vast volcanic crater, capturing its majestic and terrifying presence. The crater's jagged rims frame a deep, nearly black abyss, surrounded by a gray rock mass riddled with cracks and gullies. Suddenly, the center of the crater violently rips open, and a stream of searing white magma erupts explosively. Like a vast, luminous fountain, it shoots skyward, hurling countless molten spheres into the air. With a relentless, almost overexposed brightness, the viscous lava spills from the crater in streams of brilliant white, slowly but surely carving new paths across the vast, desolate, dark, ash-covered slopes. Simultaneously, a massive, billowing column of smoke and ash rises, its intricate gray textures creating a stark visual contrast with the fiery magma below. The entire scene is completely illuminated by the light of the magma itself, casting violently dancing shadows and creating a primal, apocalyptic atmosphere full of the irresistible and awe-inspiring power of nature.", + "seed": 11025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_1/index_1154_seed_11025_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0940266906749457, + "vlm_logprob": -0.0940266906749457, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0940266906749457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38855448365211487 + ], + "opposite_logprob_tokens": [ + -1.2635544538497925 + ], + "logprob_reward": -0.38855448365211487, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.38855448365211487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24852274358272552, + 0.07777908444404602, + 0.09816709905862808, + 0.16277562081813812, + 0.05974061042070389, + 0.040491800755262375, + 0.05437111854553223, + 0.04350446164608002, + 0.06925635039806366, + 0.14539110660552979 + ], + "credit_map": [ + 0.024045206606388092, + 0.033906254917383194, + 0.019456839188933372, + 0.026647966355085373, + 0.035076454281806946, + 0.02126740664243698, + 0.04933357983827591, + 0.0387890450656414, + 0.009111367166042328, + 0.007577496115118265, + 0.007516943383961916, + 0.010151665657758713, + 0.00996009074151516, + 0.010599188506603241, + 0.01202473882585764, + 0.010837594047188759, + 0.01612357795238495, + 0.009340663440525532, + 0.02013619989156723, + 0.008595519699156284, + 0.011522846296429634, + 0.011176767759025097, + 0.006918115075677633, + 0.014353400096297264, + 0.026777714490890503, + 0.023409204557538033, + 0.015776311978697777, + 0.017273806035518646, + 0.02080177702009678, + 0.028288496658205986, + 0.011767834424972534, + 0.018680477514863014, + 0.01112562045454979, + 0.0036644788924604654, + 0.007651142776012421, + 0.00975585076957941, + 0.004926974885165691, + 0.00984103512018919, + 0.006403209641575813, + 0.0063722999766469, + 0.007185483351349831, + 0.003984021954238415, + 0.004797653760761023, + 0.005400936119258404, + 0.003865748643875122, + 0.009016139432787895, + 0.002707101171836257, + 0.0035347146913409233, + 0.007873155176639557, + 0.004760787356644869, + 0.008424945175647736, + 0.009846989065408707, + 0.004900991450995207, + 0.008383168838918209, + 0.006097485311329365, + 0.004083598032593727, + 0.003959848079830408, + 0.005200834479182959, + 0.003956618253141642, + 0.008146370761096478, + 0.005607973784208298, + 0.003939926624298096, + 0.007387103978544474, + 0.00530578289180994, + 0.006443999242037535, + 0.011311921291053295, + 0.0037009709049016237, + 0.010880369693040848, + 0.013333948329091072, + 0.004514495376497507, + 0.014840708114206791, + 0.00422994140535593, + 0.006916474550962448, + 0.04860684275627136, + 0.007354436442255974, + 0.012801128439605236, + 0.04465491324663162, + 0.009523804299533367, + 0.009030655957758427, + 0.00650285929441452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.38855448365211487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00325388228520751 + ], + "opposite_logprob_tokens": [ + -5.753253936767578 + ], + "logprob_reward": -0.00325388228520751, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00325388228520751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15582142770290375, + 0.11662676185369492, + 0.11917388439178467, + 0.09862058609724045, + 0.032339178025722504, + 0.02844604104757309, + 0.03035580739378929, + 0.06402691453695297, + 0.03565697744488716, + 0.3189324140548706 + ], + "credit_map": [ + 0.021445292979478836, + 0.02049499750137329, + 0.014866397716104984, + 0.017551107332110405, + 0.020972708240151405, + 0.019227037206292152, + 0.019996607676148415, + 0.021267283707857132, + 0.014062881469726562, + 0.014193018898367882, + 0.012333196587860584, + 0.01516047865152359, + 0.014085809700191021, + 0.011030090041458607, + 0.019571471959352493, + 0.0161898136138916, + 0.020017875358462334, + 0.007340496871620417, + 0.029414484277367592, + 0.009426740929484367, + 0.010328392498195171, + 0.01428278163075447, + 0.006435276009142399, + 0.021927833557128906, + 0.014123823493719101, + 0.01330332551151514, + 0.007426545489579439, + 0.010798810981214046, + 0.01567268557846546, + 0.016493694856762886, + 0.011242857202887535, + 0.009558841586112976, + 0.00595936318859458, + 0.0027301700320094824, + 0.0032196762040257454, + 0.00450778566300869, + 0.0038985852152109146, + 0.00539447832852602, + 0.0025396309792995453, + 0.0040894863195717335, + 0.005122255999594927, + 0.003211574163287878, + 0.0035636499524116516, + 0.003266260726377368, + 0.0023599539417773485, + 0.005318489391356707, + 0.00305164884775877, + 0.002552209421992302, + 0.00491426233202219, + 0.002986398059874773, + 0.004619510378688574, + 0.003719126107171178, + 0.0026958046946674585, + 0.006503152195364237, + 0.0025933273136615753, + 0.002324227476492524, + 0.006145137827843428, + 0.009366452693939209, + 0.006583135575056076, + 0.010472146794199944, + 0.0098331393674016, + 0.005504129454493523, + 0.010541011579334736, + 0.005581758450716734, + 0.0038802397903054953, + 0.005616007838398218, + 0.0023524716962128878, + 0.005861465819180012, + 0.0069536687806248665, + 0.002728185150772333, + 0.006462278310209513, + 0.0018026609905064106, + 0.011254848912358284, + 0.12931671738624573, + 0.010179322212934494, + 0.011058986186981201, + 0.12405545264482498, + 0.011640678159892559, + 0.013769599609076977, + 0.007656833156943321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.00325388228520751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004649192094803 + ], + "opposite_logprob_tokens": [ + -2.6950464248657227 + ], + "logprob_reward": -0.07004649192094803, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004649192094803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10498207807540894, + 0.05829339101910591, + 0.07208579778671265, + 0.09287254512310028, + 0.026155870407819748, + 0.05063777044415474, + 0.052261680364608765, + 0.035773515701293945, + 0.04193294420838356, + 0.4650043547153473 + ], + "credit_map": [ + 0.00975774135440588, + 0.014494436793029308, + 0.00800693966448307, + 0.00907075684517622, + 0.01492859236896038, + 0.013028038665652275, + 0.020174751058220863, + 0.015520825982093811, + 0.00861204694956541, + 0.006042992230504751, + 0.005752586293965578, + 0.008079363964498043, + 0.006598202977329493, + 0.007903476245701313, + 0.007227702997624874, + 0.008077019825577736, + 0.010000326670706272, + 0.005783413536846638, + 0.010108965449035168, + 0.006971505470573902, + 0.010603531263768673, + 0.007471834309399128, + 0.009707370772957802, + 0.011438854970037937, + 0.013775368221104145, + 0.008610496297478676, + 0.009619489312171936, + 0.008882851339876652, + 0.01946195960044861, + 0.010841146111488342, + 0.012534328736364841, + 0.009146898053586483, + 0.003981712274253368, + 0.0029046828858554363, + 0.0023361481726169586, + 0.003828061278909445, + 0.0031021377071738243, + 0.004671361763030291, + 0.0024752048775553703, + 0.002856561215594411, + 0.0112973153591156, + 0.0054467166773974895, + 0.005706439726054668, + 0.00610205065459013, + 0.004003643523901701, + 0.008113963529467583, + 0.0044852616265416145, + 0.00548237981274724, + 0.007480457425117493, + 0.006576972547918558, + 0.007865117862820625, + 0.008060487918555737, + 0.005286086350679398, + 0.0070521030575037, + 0.0060772281140089035, + 0.003863227553665638, + 0.003572839545086026, + 0.003933944273740053, + 0.005021057557314634, + 0.006533712614327669, + 0.004150843247771263, + 0.0027545427437871695, + 0.0053382995538413525, + 0.004468272440135479, + 0.005512676201760769, + 0.006303205620497465, + 0.0026594374794512987, + 0.0052023171447217464, + 0.008484918624162674, + 0.0028732060454785824, + 0.007574958726763725, + 0.003322224598377943, + 0.015583240427076817, + 0.18010206520557404, + 0.015940561890602112, + 0.020235562697052956, + 0.17133228480815887, + 0.021109934896230698, + 0.024530652910470963, + 0.016170071437954903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07004649192094803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003630714723840356 + ], + "opposite_logprob_tokens": [ + -5.628630638122559 + ], + "logprob_reward": -0.003630714723840356, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003630714723840356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15413978695869446, + 0.03588205203413963, + 0.03581976518034935, + 0.27658188343048096, + 0.05808047577738762, + 0.043588485568761826, + 0.08919714391231537, + 0.16635654866695404, + 0.06529248505830765, + 0.0750613585114479 + ], + "credit_map": [ + 0.01657828502357006, + 0.021070245653390884, + 0.012027112767100334, + 0.016713377088308334, + 0.02340248040854931, + 0.01436240877956152, + 0.02574247121810913, + 0.02424340695142746, + 0.004351073410362005, + 0.0044227479957044125, + 0.0036617815494537354, + 0.004026617389172316, + 0.0045163934119045734, + 0.0029905596747994423, + 0.007083910517394543, + 0.004828969016671181, + 0.005973692052066326, + 0.002956429962068796, + 0.005804537795484066, + 0.003374146530404687, + 0.004671239759773016, + 0.003930136561393738, + 0.004506015684455633, + 0.004603566136211157, + 0.03753043711185455, + 0.029274705797433853, + 0.023499714210629463, + 0.028750652447342873, + 0.05625394359230995, + 0.035206787288188934, + 0.0391688197851181, + 0.02689683996140957, + 0.014998995698988438, + 0.004239980131387711, + 0.008616389706730843, + 0.006443486548960209, + 0.004424501676112413, + 0.007735728286206722, + 0.005988173186779022, + 0.005633220076560974, + 0.00823313370347023, + 0.006395982578396797, + 0.004543795716017485, + 0.004701907746493816, + 0.003977233078330755, + 0.006108639296144247, + 0.005147572141140699, + 0.004480219446122646, + 0.015311038121581078, + 0.00795330386608839, + 0.015430386178195477, + 0.009592008776962757, + 0.007751026656478643, + 0.018015678972005844, + 0.008244119584560394, + 0.006899585947394371, + 0.016502929851412773, + 0.023820750415325165, + 0.015825890004634857, + 0.03144346922636032, + 0.02574806474149227, + 0.01428397186100483, + 0.025327909737825394, + 0.013403565622866154, + 0.006544989068061113, + 0.008292250335216522, + 0.003778004553169012, + 0.008663471788167953, + 0.01053108274936676, + 0.005898594856262207, + 0.015363190323114395, + 0.006220899522304535, + 0.006399772595614195, + 0.016575781628489494, + 0.004494993481785059, + 0.01052170991897583, + 0.014615748077630997, + 0.008094459772109985, + 0.008058000355958939, + 0.0063008880242705345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.003630714723840356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone view looking down into a volcanic crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in black-and-white with a gray rock mass and jagged crater rims?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly inch toward the center of the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does white magma erupt explosively from the center, shooting skyward as a luminous fountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the magma's light, creating overexposed brightness and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647880792617798 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647880792617798, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647880792617798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11052969843149185, + 0.0714571550488472, + 0.04085364565253258, + 0.12610512971878052, + 0.04462938383221626, + 0.06131195276975632, + 0.04571199789643288, + 0.08612848073244095, + 0.0425015389919281, + 0.37077105045318604 + ], + "credit_map": [ + 0.01477140374481678, + 0.014561569318175316, + 0.01037939265370369, + 0.011771350167691708, + 0.01551846507936716, + 0.013056520372629166, + 0.012878237292170525, + 0.01759275235235691, + 0.009191394783556461, + 0.007959390990436077, + 0.007857188582420349, + 0.008456719107925892, + 0.008372295647859573, + 0.011604680679738522, + 0.008621868677437305, + 0.009393611922860146, + 0.006656981073319912, + 0.0036657797172665596, + 0.005268780514597893, + 0.0040237512439489365, + 0.006028646137565374, + 0.004808198660612106, + 0.005899242125451565, + 0.0045022680424153805, + 0.0180360060185194, + 0.013432453386485577, + 0.010494140908122063, + 0.01363990269601345, + 0.024727756157517433, + 0.014839672483503819, + 0.01981702260673046, + 0.011118187569081783, + 0.00887911207973957, + 0.004358153324574232, + 0.004767871927469969, + 0.00596002908423543, + 0.004529599100351334, + 0.006095043383538723, + 0.004785649478435516, + 0.00525392172858119, + 0.00842253863811493, + 0.008923404850065708, + 0.007320871111005545, + 0.00919425580650568, + 0.006011004094034433, + 0.008814485743641853, + 0.007264568004757166, + 0.005360829643905163, + 0.008630525320768356, + 0.005196852143853903, + 0.0070990934036672115, + 0.0052392748184502125, + 0.005144891794770956, + 0.00718345632776618, + 0.004397835582494736, + 0.002820067573338747, + 0.009162659756839275, + 0.010933036915957928, + 0.007968263700604439, + 0.015709562227129936, + 0.01235700398683548, + 0.007594836410135031, + 0.014781425707042217, + 0.007621689699590206, + 0.004486993886530399, + 0.007119815796613693, + 0.003065957687795162, + 0.005620298907160759, + 0.0074255927465856075, + 0.002603133674710989, + 0.009146466851234436, + 0.0030332820024341345, + 0.012271587736904621, + 0.15135090053081512, + 0.011771118268370628, + 0.012966601178050041, + 0.14411655068397522, + 0.013518175110220909, + 0.015461700968444347, + 0.009314417839050293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.004647880792617798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_tensor([9265, 9265])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_tensor([9265, 9265])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90e60e21730e86ab16b172a436e268541f7451b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_tensor([9265, 9265])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In Steve McCurry's signature rich color style, a slow-motion medium shot, captured with high-speed cinematography, precisely freezes a moment brimming with dramatic tension. A warm, intense afternoon light slants down, illuminating a weathered, mottled wall. The surface is a rich interplay of deep ochre yellow and faded turquoise blue. The paint, eroded by age, has flaked off, revealing the rough plaster beneath. Before this storied wall, a graceful swallow, startled by an unknown source, suddenly takes flight, its body tensed in the air. Caught in extreme slow motion, its back, shimmering with metallic blue-green feathers, stands in stark contrast to its soft, creamy white belly. Its wings sweep upward with breathtaking force, fully extended. The outline of each feather is clearly visible in the air, and the dust and wall debris that rise up create a fleeting halo of light around it. The color saturation of the entire picture is extremely high, and the light makes the warm tones of the wall particularly mellow, forming a distinct and harmonious visual impact with the cool luster of the swallow's feathers, creating a strong atmosphere of vitality bursting out in chaos and capturing eternity in the fleeting moment.", + "seed": 19477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19477_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0095691459486261, + "vlm_logprob": -0.0095691459486261, + "vlm_yes_no_margin": 5.75, + "vlm_token_credit": -0.0095691459486261, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0016689670737832785 + ], + "opposite_logprob_tokens": [ + -7.626668930053711 + ], + "logprob_reward": -0.0016689670737832785, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0016689670737832785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11369369924068451, + 0.047397445887327194, + 0.08009301126003265, + 0.09111694991588593, + 0.04901923984289169, + 0.3369871973991394, + 0.033416856080293655, + 0.08949301391839981, + 0.10303039103746414, + 0.0557522289454937 + ], + "credit_map": [ + 0.010564262047410011, + 0.012500077486038208, + 0.01761269010603428, + 0.019555239006876945, + 0.014019106514751911, + 0.013799423351883888, + 0.01491578109562397, + 0.010727114044129848, + 0.00412767706438899, + 0.006922710686922073, + 0.0031084741931408644, + 0.005392934661358595, + 0.005530716851353645, + 0.010446466505527496, + 0.006016113795340061, + 0.005852350499480963, + 0.008597824722528458, + 0.00825897790491581, + 0.009338968433439732, + 0.008174615912139416, + 0.010650523006916046, + 0.015062220394611359, + 0.010449554771184921, + 0.009560314007103443, + 0.007487605791538954, + 0.008538978174328804, + 0.010107691399753094, + 0.02547347918152809, + 0.008288593031466007, + 0.012639685533940792, + 0.009519084356725216, + 0.009061831049621105, + 0.005632367916405201, + 0.004877332132309675, + 0.005448572803288698, + 0.005779394414275885, + 0.005146447103470564, + 0.007005488965660334, + 0.005723302252590656, + 0.00940633099526167, + 0.029072046279907227, + 0.03127526119351387, + 0.040716398507356644, + 0.03916138410568237, + 0.049893759191036224, + 0.04008117690682411, + 0.043983619660139084, + 0.06280352920293808, + 0.0030942722223699093, + 0.004036659840494394, + 0.005108331795781851, + 0.003702949732542038, + 0.004164922051131725, + 0.004402829334139824, + 0.005321963224560022, + 0.003584926947951317, + 0.00826149620115757, + 0.017091089859604836, + 0.009364706464111805, + 0.00989616196602583, + 0.01490111369639635, + 0.012805634178221226, + 0.009933671914041042, + 0.007239127531647682, + 0.01577364094555378, + 0.012248102575540543, + 0.02455429546535015, + 0.00857136957347393, + 0.011514361016452312, + 0.01278084609657526, + 0.009440777823328972, + 0.008146993815898895, + 0.005597816780209541, + 0.007259683683514595, + 0.010981331579387188, + 0.003963025286793709, + 0.007436587009578943, + 0.00797799788415432, + 0.0065509239211678505, + 0.005984861869364977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0016689670737832785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13726888597011566, + 0.02446834370493889, + 0.015017591416835785, + 0.016511758789420128, + 0.08074212074279785, + 0.02237960696220398, + 0.17652712762355804, + 0.07301922142505646, + 0.41892778873443604, + 0.03513756021857262 + ], + "credit_map": [ + 0.011031366884708405, + 0.016715656965970993, + 0.017330951988697052, + 0.01443330105394125, + 0.01715170405805111, + 0.02350343018770218, + 0.0163761954754591, + 0.02072627656161785, + 0.0023875688202679157, + 0.0025036283768713474, + 0.0017063611885532737, + 0.002261033048853278, + 0.0032237928826361895, + 0.004945769906044006, + 0.0032324399799108505, + 0.0042077479884028435, + 0.0015649948036298156, + 0.0014854128239676356, + 0.002226215321570635, + 0.0016598938964307308, + 0.002020446350798011, + 0.001918058842420578, + 0.002065531210973859, + 0.002077037701383233, + 0.0016224767314270139, + 0.0017022717511281371, + 0.0013412146363407373, + 0.003802178194746375, + 0.0019320679130032659, + 0.002222847891971469, + 0.0018307280261069536, + 0.0020579739939421415, + 0.006275472231209278, + 0.006497533991932869, + 0.0068131135776638985, + 0.004708173684775829, + 0.016408102586865425, + 0.019955871626734734, + 0.008143533021211624, + 0.011940316297113895, + 0.0032230010256171227, + 0.0021389415487647057, + 0.003131998237222433, + 0.002169138053432107, + 0.0026290963869541883, + 0.002809664234519005, + 0.00247055827639997, + 0.0038072087336331606, + 0.01483219675719738, + 0.021372206509113312, + 0.017199022695422173, + 0.017786892130970955, + 0.025421738624572754, + 0.021248269826173782, + 0.02709897607564926, + 0.03156781941652298, + 0.00724714295938611, + 0.012301988899707794, + 0.006733219139277935, + 0.007762105669826269, + 0.01358699705451727, + 0.011045712046325207, + 0.0069879875518381596, + 0.007354074157774448, + 0.038361892104148865, + 0.07164817303419113, + 0.04558579623699188, + 0.04504096880555153, + 0.07273315638303757, + 0.06996459513902664, + 0.044048964977264404, + 0.03154426068067551, + 0.0029451828449964523, + 0.0031885793432593346, + 0.005435631610453129, + 0.0034060480538755655, + 0.0032020036596804857, + 0.005545952823013067, + 0.004900826141238213, + 0.00651333574205637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017115719383582473 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017115719383582473, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017115719383582473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05873074382543564, + 0.039870310574769974, + 0.029769612476229668, + 0.07331187278032303, + 0.1749320775270462, + 0.39466845989227295, + 0.0398816354572773, + 0.05123307183384895, + 0.0628843829035759, + 0.07471787184476852 + ], + "credit_map": [ + 0.005693261977285147, + 0.006964963395148516, + 0.008732878603041172, + 0.007831712253391743, + 0.007572873495519161, + 0.007278928533196449, + 0.007569490000605583, + 0.007086632307618856, + 0.004252457059919834, + 0.0046126642264425755, + 0.004099608864635229, + 0.005338061600923538, + 0.005155207123607397, + 0.008316446095705032, + 0.003898525843396783, + 0.004197335336357355, + 0.0027613781858235598, + 0.0033633820712566376, + 0.003955837804824114, + 0.0028821323066949844, + 0.0037664659321308136, + 0.00686518894508481, + 0.00301660830155015, + 0.00315862032584846, + 0.005835088435560465, + 0.006831910461187363, + 0.006542399991303682, + 0.01830539107322693, + 0.011357511393725872, + 0.012184466235339642, + 0.0067247371189296246, + 0.005530360154807568, + 0.013188406825065613, + 0.01355310995131731, + 0.017030898481607437, + 0.01402437873184681, + 0.03381207957863808, + 0.043879032135009766, + 0.015953494235873222, + 0.023490671068429947, + 0.033003609627485275, + 0.039314668625593185, + 0.048504456877708435, + 0.047292180359363556, + 0.057608019560575485, + 0.04678649082779884, + 0.05246627703309059, + 0.06969273835420609, + 0.0036976803094148636, + 0.006963996682316065, + 0.003647979348897934, + 0.005302684381604195, + 0.005897645838558674, + 0.005841617006808519, + 0.004049500450491905, + 0.00448053190484643, + 0.00472545949742198, + 0.010339181870222092, + 0.004530016798526049, + 0.005690015386790037, + 0.009621774777770042, + 0.008109460584819317, + 0.004801140166819096, + 0.003416022751480341, + 0.0051001934334635735, + 0.009545564651489258, + 0.006838629953563213, + 0.006406777538359165, + 0.01374195609241724, + 0.010379833169281483, + 0.006418343633413315, + 0.00445308443158865, + 0.008694810792803764, + 0.008597763255238533, + 0.015135488472878933, + 0.0045939297415316105, + 0.007731168065220118, + 0.015339885838329792, + 0.006935213226824999, + 0.007689614314585924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.0017115719383582473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031881716568022966 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.0031881716568022966, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031881716568022966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0778304785490036, + 0.0506553128361702, + 0.03526047244668007, + 0.18294712901115417, + 0.13409528136253357, + 0.25008854269981384, + 0.030836496502161026, + 0.06263422220945358, + 0.12687087059020996, + 0.04878116771578789 + ], + "credit_map": [ + 0.006968559697270393, + 0.008280533365905285, + 0.011579730547964573, + 0.011470240540802479, + 0.009406395256519318, + 0.01267714612185955, + 0.008882435038685799, + 0.00856543704867363, + 0.004657757468521595, + 0.004655088298022747, + 0.006234853062778711, + 0.008194314315915108, + 0.006060355808585882, + 0.009437805972993374, + 0.005632579326629639, + 0.0057825553230941296, + 0.0040741716511547565, + 0.004240245092660189, + 0.0028680013492703438, + 0.003961111418902874, + 0.004022432025521994, + 0.008876667357981205, + 0.004111059941351414, + 0.0031067836098372936, + 0.014103928580880165, + 0.01628408022224903, + 0.021527260541915894, + 0.06075432151556015, + 0.01742558367550373, + 0.028712617233395576, + 0.012772765941917896, + 0.011366567574441433, + 0.013694687746465206, + 0.012149112299084663, + 0.012542559765279293, + 0.007631639484316111, + 0.02470622956752777, + 0.030747588723897934, + 0.01199901569634676, + 0.020624440163373947, + 0.02202531322836876, + 0.022511422634124756, + 0.0290228221565485, + 0.028854437172412872, + 0.04306873679161072, + 0.03110182099044323, + 0.034693676978349686, + 0.03881031647324562, + 0.0023955833166837692, + 0.005411030258983374, + 0.003813762916252017, + 0.0036915361415594816, + 0.005283578298985958, + 0.003958005458116531, + 0.0029601603746414185, + 0.00332283996976912, + 0.006609967909753323, + 0.011859522201120853, + 0.006180320400744677, + 0.006696722935885191, + 0.009787370450794697, + 0.008759303949773312, + 0.006770626176148653, + 0.005970389116555452, + 0.03367193415760994, + 0.007752149365842342, + 0.053954798728227615, + 0.006316662300378084, + 0.004532113205641508, + 0.01172454934567213, + 0.004159416072070599, + 0.004759249743074179, + 0.004615345038473606, + 0.00578159699216485, + 0.005978499073535204, + 0.003931361250579357, + 0.007709961384534836, + 0.009951302781701088, + 0.0053506153635680676, + 0.005462488159537315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.0031881716568022966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808433189988136 + ], + "opposite_logprob_tokens": [ + -3.2880842685699463 + ], + "logprob_reward": -0.03808433189988136, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808433189988136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10483641922473907, + 0.06004706770181656, + 0.10874968767166138, + 0.06297364830970764, + 0.1128956526517868, + 0.12544766068458557, + 0.039937857538461685, + 0.08846214413642883, + 0.15897199511528015, + 0.13767792284488678 + ], + "credit_map": [ + 0.009019461460411549, + 0.011749764904379845, + 0.012411307543516159, + 0.020130831748247147, + 0.013765266165137291, + 0.013586368411779404, + 0.01180239301174879, + 0.012371012941002846, + 0.00747745530679822, + 0.008053670637309551, + 0.005211203824728727, + 0.007371437735855579, + 0.007357013411819935, + 0.010683546774089336, + 0.006695852614939213, + 0.007196884136646986, + 0.012675277888774872, + 0.011667679063975811, + 0.006988913752138615, + 0.00951776746660471, + 0.014960936270654202, + 0.026019761338829994, + 0.014224003069102764, + 0.012695342302322388, + 0.004801924340426922, + 0.00602068891748786, + 0.006410868372768164, + 0.01375892385840416, + 0.00954369269311428, + 0.008538180962204933, + 0.007166546303778887, + 0.0067328219301998615, + 0.014368281699717045, + 0.011004287749528885, + 0.010956941172480583, + 0.00737868994474411, + 0.015181124210357666, + 0.01832953281700611, + 0.014512150548398495, + 0.021164648234844208, + 0.012064515613019466, + 0.01364302821457386, + 0.014248247258365154, + 0.014313549734652042, + 0.014794698916375637, + 0.015614461153745651, + 0.016640160232782364, + 0.024128993973135948, + 0.004658389370888472, + 0.007092311047017574, + 0.005153736099600792, + 0.005173444747924805, + 0.005471307784318924, + 0.005636794958263636, + 0.003920460119843483, + 0.0028314136434346437, + 0.010426337830722332, + 0.008710789494216442, + 0.018628597259521484, + 0.010562215931713581, + 0.007975304499268532, + 0.013699037954211235, + 0.007889371365308762, + 0.010570481419563293, + 0.039048995822668076, + 0.009998152032494545, + 0.06193598732352257, + 0.008899635635316372, + 0.008914606645703316, + 0.018034325912594795, + 0.0061307065188884735, + 0.006009566131979227, + 0.01576245203614235, + 0.014625795185565948, + 0.025645099580287933, + 0.009866981767117977, + 0.015831898897886276, + 0.027033399790525436, + 0.014569884166121483, + 0.01434238813817501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.03808433189988136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Steve McCurry's signature rich color style, a slow-motion medium shot, captured with high-speed cinematography, precisely freezes a moment brimming with dramatic tension. A warm, intense afternoon light slants down, illuminating a weathered, mottled wall. The surface is a rich interplay of deep ochre yellow and faded turquoise blue. The paint, eroded by age, has flaked off, revealing the rough plaster beneath. Before this storied wall, a graceful swallow, startled by an unknown source, suddenly takes flight, its body tensed in the air. Caught in extreme slow motion, its back, shimmering with metallic blue-green feathers, stands in stark contrast to its soft, creamy white belly. Its wings sweep upward with breathtaking force, fully extended. The outline of each feather is clearly visible in the air, and the dust and wall debris that rise up create a fleeting halo of light around it. The color saturation of the entire picture is extremely high, and the light makes the warm tones of the wall particularly mellow, forming a distinct and harmonious visual impact with the cool luster of the swallow's feathers, creating a strong atmosphere of vitality bursting out in chaos and capturing eternity in the fleeting moment.", + "seed": 19478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19478_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010421984968706966, + "vlm_logprob": -0.010421984968706966, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.010421984968706966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0020507988519966602 + ], + "opposite_logprob_tokens": [ + -7.252050876617432 + ], + "logprob_reward": -0.0020507988519966602, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0020507988519966602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07752250134944916, + 0.03662097826600075, + 0.0673808604478836, + 0.16926491260528564, + 0.08037669211626053, + 0.02614440582692623, + 0.26272305846214294, + 0.019559815526008606, + 0.07973594963550568, + 0.18067088723182678 + ], + "credit_map": [ + 0.008675468154251575, + 0.012409690767526627, + 0.008809985592961311, + 0.004892678000032902, + 0.011890178546309471, + 0.012452746741473675, + 0.010118993930518627, + 0.008272764272987843, + 0.004100805148482323, + 0.004560451488941908, + 0.00580111937597394, + 0.0033295874018222094, + 0.00583491800352931, + 0.0048506539314985275, + 0.005181423854082823, + 0.0029620209243148565, + 0.010153744369745255, + 0.005072258412837982, + 0.01929592899978161, + 0.0055052549578249454, + 0.006998381111770868, + 0.00801155623048544, + 0.0052502769976854324, + 0.007093454245477915, + 0.04919916018843651, + 0.016553232446312904, + 0.049942366778850555, + 0.015371120534837246, + 0.008273775689303875, + 0.013007071800529957, + 0.008221910335123539, + 0.008696281351149082, + 0.014517734758555889, + 0.0071707977913320065, + 0.03251637890934944, + 0.0059665413573384285, + 0.002575522754341364, + 0.003369510406628251, + 0.0034181526862084866, + 0.01084205787628889, + 0.0038428674452006817, + 0.0034842747263610363, + 0.004045384004712105, + 0.0036518373526632786, + 0.0030396683141589165, + 0.0032629924826323986, + 0.0021924946922808886, + 0.0026248865760862827, + 0.0925540030002594, + 0.01712859608232975, + 0.10768695920705795, + 0.01678551733493805, + 0.0043782927095890045, + 0.007350017782300711, + 0.007201789878308773, + 0.009637875482439995, + 0.0032334085553884506, + 0.0028989638667553663, + 0.002492172410711646, + 0.0026641814038157463, + 0.001850712113082409, + 0.002070101909339428, + 0.0022035520523786545, + 0.002146722748875618, + 0.016010915860533714, + 0.008696364238858223, + 0.023972200229763985, + 0.011277277953922749, + 0.003947013057768345, + 0.00500112259760499, + 0.006146974861621857, + 0.004684078507125378, + 0.047703035175800323, + 0.029492130503058434, + 0.021430911496281624, + 0.02296759933233261, + 0.009895218536257744, + 0.008177570067346096, + 0.013402934186160564, + 0.027601497247815132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.0020507988519966602 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049824535846710205, + 0.04049840569496155, + 0.22204452753067017, + 0.2802048623561859, + 0.14690449833869934, + 0.07561757415533066, + 0.03120742179453373, + 0.011595581658184528, + 0.09119545668363571, + 0.050907135009765625 + ], + "credit_map": [ + 0.004705507308244705, + 0.006426693871617317, + 0.005645342171192169, + 0.0037111383862793446, + 0.0063546281307935715, + 0.007657488342374563, + 0.006914424244314432, + 0.008409318514168262, + 0.005195569712668657, + 0.007114843465387821, + 0.005463222973048687, + 0.0031867199577391148, + 0.007292689755558968, + 0.005163750611245632, + 0.003885875688865781, + 0.0031957379542291164, + 0.04020380973815918, + 0.021107986569404602, + 0.0613706149160862, + 0.016165541484951973, + 0.02040150761604309, + 0.03200859948992729, + 0.015255135484039783, + 0.015531355515122414, + 0.08516118675470352, + 0.030380524694919586, + 0.08222059160470963, + 0.022214172407984734, + 0.014447197318077087, + 0.017547335475683212, + 0.01443587988615036, + 0.013798005878925323, + 0.03107902593910694, + 0.011976060457527637, + 0.06367972493171692, + 0.009836922399699688, + 0.0025501747149974108, + 0.0038704832550138235, + 0.00532586220651865, + 0.018586253747344017, + 0.011826143600046635, + 0.00885509792715311, + 0.011755960993468761, + 0.011364021338522434, + 0.006449915934354067, + 0.011249681934714317, + 0.0069939931854605675, + 0.0071227713488042355, + 0.009926596656441689, + 0.0025281875859946012, + 0.011139030568301678, + 0.002282693749293685, + 0.0011729864636436105, + 0.0011924561113119125, + 0.0012812595814466476, + 0.0016842132899910212, + 0.002866077935323119, + 0.0011938546085730195, + 0.0034941122867166996, + 0.001202378305606544, + 0.000740712508559227, + 0.0005220890161581337, + 0.000767185993026942, + 0.0008091725176200271, + 0.011851799674332142, + 0.01155423279851675, + 0.024262042716145515, + 0.012283100746572018, + 0.005440459121018648, + 0.006969335488975048, + 0.008439121767878532, + 0.010395369492471218, + 0.017696088179945946, + 0.008794405497610569, + 0.004574785474687815, + 0.002654365263879299, + 0.0024998413864523172, + 0.002409122185781598, + 0.003584420308470726, + 0.008694113232195377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019375376868993044 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019375376868993044, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019375376868993044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019441360607743263, + 0.03441793844103813, + 0.04479667916893959, + 0.026657937094569206, + 0.11836952716112137, + 0.014078324660658836, + 0.10982195287942886, + 0.013080933131277561, + 0.025218743830919266, + 0.5941165685653687 + ], + "credit_map": [ + 0.00214609713293612, + 0.004381804261356592, + 0.002321097766980529, + 0.0014700341271236539, + 0.0035374839790165424, + 0.0015056021511554718, + 0.00234036217443645, + 0.001738879014737904, + 0.004332341253757477, + 0.007671593688428402, + 0.004327799659222364, + 0.0033762771636247635, + 0.005341432057321072, + 0.002882304834201932, + 0.003957257140427828, + 0.002528934506699443, + 0.007617192808538675, + 0.004741264972835779, + 0.011404052376747131, + 0.003450101939961314, + 0.004237794317305088, + 0.00623381556943059, + 0.0034391451627016068, + 0.003673313884064555, + 0.0062355762347579, + 0.0031221576500684023, + 0.006222455762326717, + 0.002894570818170905, + 0.002747755730524659, + 0.0017351810820400715, + 0.0017531239427626133, + 0.0019471141276881099, + 0.02480047382414341, + 0.009669187478721142, + 0.05044486001133919, + 0.0075312452390789986, + 0.002943499945104122, + 0.0037828045897185802, + 0.005419638939201832, + 0.013777827844023705, + 0.0020551278721541166, + 0.0020932552870362997, + 0.0020930625032633543, + 0.0021513435058295727, + 0.001200400642119348, + 0.0018771501490846276, + 0.0012670839205384254, + 0.0013409012462943792, + 0.036989420652389526, + 0.00696669751778245, + 0.047448594123125076, + 0.007285425439476967, + 0.0022046815138310194, + 0.0022373239044100046, + 0.0021968933288007975, + 0.004492915701121092, + 0.0023412101436406374, + 0.0015487216878682375, + 0.0029140322003513575, + 0.0016117554623633623, + 0.0010338608408346772, + 0.0009747811127454042, + 0.0014491743640974164, + 0.0012073979014530778, + 0.0045903585851192474, + 0.003022334538400173, + 0.008039817214012146, + 0.0027214556466788054, + 0.0015554664423689246, + 0.001618751441128552, + 0.0015495122643187642, + 0.002121047582477331, + 0.2018837332725525, + 0.09158232063055038, + 0.06915504485368729, + 0.04063206538558006, + 0.029746949672698975, + 0.026744619011878967, + 0.03961316496133804, + 0.09475868940353394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.0019375376868993044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090751986950636 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090751986950636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090751986950636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03435888886451721, + 0.03714980185031891, + 0.10343464463949203, + 0.016845718026161194, + 0.12186062335968018, + 0.03780307620763779, + 0.48611801862716675, + 0.03426465392112732, + 0.033345118165016174, + 0.09481952339410782 + ], + "credit_map": [ + 0.004070084076374769, + 0.007819708436727524, + 0.00379977747797966, + 0.0031919211614876986, + 0.005654231645166874, + 0.0026650845538824797, + 0.0038280594162642956, + 0.0033300218638032675, + 0.004753546789288521, + 0.005562040489166975, + 0.006134882569313049, + 0.003983959089964628, + 0.005609285086393356, + 0.004672667011618614, + 0.00384342554025352, + 0.002589993178844452, + 0.019412077963352203, + 0.01015097089111805, + 0.02647491544485092, + 0.008017187938094139, + 0.009356245398521423, + 0.014161628670990467, + 0.008716640993952751, + 0.0071449680253863335, + 0.0022870732937008142, + 0.0032946188002824783, + 0.003368941368535161, + 0.0021830350160598755, + 0.0018035719403997064, + 0.0014860769733786583, + 0.0011828868882730603, + 0.0012395121157169342, + 0.02523791417479515, + 0.010169045068323612, + 0.052997685968875885, + 0.007325248327106237, + 0.0023087572772055864, + 0.003174516838043928, + 0.004559330642223358, + 0.016088111326098442, + 0.006673688068985939, + 0.005227706395089626, + 0.006172564812004566, + 0.0046256305649876595, + 0.004274956416338682, + 0.004389755893498659, + 0.0033400682732462883, + 0.0030987027566879988, + 0.17017649114131927, + 0.03047860600054264, + 0.21055912971496582, + 0.032166317105293274, + 0.006634812336415052, + 0.009618547745049, + 0.009590994566679, + 0.01689307764172554, + 0.004869136027991772, + 0.004762999713420868, + 0.005913303699344397, + 0.0042692213319242, + 0.003426648210734129, + 0.0036949324421584606, + 0.003916111309081316, + 0.0034122997894883156, + 0.007363328710198402, + 0.0037168399430811405, + 0.011104383505880833, + 0.003021319629624486, + 0.002480102935805917, + 0.0019224474672228098, + 0.0016740536084398627, + 0.0020626375917345285, + 0.022391241043806076, + 0.01701972261071205, + 0.012287146411836147, + 0.008908378891646862, + 0.005777965299785137, + 0.0062667918391525745, + 0.007035876624286175, + 0.015132391825318336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.004090751986950636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808249533176422 + ], + "opposite_logprob_tokens": [ + -3.2880825996398926 + ], + "logprob_reward": -0.03808249533176422, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808249533176422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08881699293851852, + 0.04626576974987984, + 0.09251680225133896, + 0.11348092555999756, + 0.07823269814252853, + 0.11833173036575317, + 0.07646528631448746, + 0.0262424536049366, + 0.03333038091659546, + 0.3263169527053833 + ], + "credit_map": [ + 0.010736040771007538, + 0.015643445774912834, + 0.007251670118421316, + 0.005818277131766081, + 0.017833568155765533, + 0.011805650778114796, + 0.012364264577627182, + 0.007364068180322647, + 0.008458258584141731, + 0.006764254067093134, + 0.008316435851156712, + 0.0029036044143140316, + 0.007986973039805889, + 0.005390747915953398, + 0.003404295304790139, + 0.0030411959160119295, + 0.0152180390432477, + 0.008735557086765766, + 0.026601361110806465, + 0.006660061422735453, + 0.008764492347836494, + 0.013624463230371475, + 0.0065918127074837685, + 0.0063210055232048035, + 0.030282314866781235, + 0.013357221148908138, + 0.03005291521549225, + 0.010612847283482552, + 0.006944975350052118, + 0.009580031037330627, + 0.006283883936703205, + 0.006366726476699114, + 0.01737944781780243, + 0.006158505566418171, + 0.02384503372013569, + 0.011172479018568993, + 0.002800963819026947, + 0.007464735768735409, + 0.004880343098193407, + 0.004531184211373329, + 0.01896069385111332, + 0.014169080182909966, + 0.020156705752015114, + 0.01762649044394493, + 0.009329133667051792, + 0.016810065135359764, + 0.010490975342690945, + 0.010788578540086746, + 0.02576421946287155, + 0.004872623831033707, + 0.03271867334842682, + 0.0050564962439239025, + 0.0014027991564944386, + 0.0019544996321201324, + 0.0017710143001750112, + 0.0029249556828290224, + 0.006267644930630922, + 0.0025016432628035545, + 0.008343163877725601, + 0.0032111462205648422, + 0.0014694652054458857, + 0.0014711122494190931, + 0.0015145448269322515, + 0.0014637329149991274, + 0.0061325738206505775, + 0.004190055653452873, + 0.011188848875463009, + 0.0030186676885932684, + 0.002458370989188552, + 0.0018714965553954244, + 0.0017517029773443937, + 0.002718662843108177, + 0.11126738041639328, + 0.04979316517710686, + 0.03871988132596016, + 0.022832905873656273, + 0.01554833073168993, + 0.01444829348474741, + 0.02207239717245102, + 0.05163458734750748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.03808249533176422 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Steve McCurry's signature rich color style, a slow-motion medium shot, captured with high-speed cinematography, precisely freezes a moment brimming with dramatic tension. A warm, intense afternoon light slants down, illuminating a weathered, mottled wall. The surface is a rich interplay of deep ochre yellow and faded turquoise blue. The paint, eroded by age, has flaked off, revealing the rough plaster beneath. Before this storied wall, a graceful swallow, startled by an unknown source, suddenly takes flight, its body tensed in the air. Caught in extreme slow motion, its back, shimmering with metallic blue-green feathers, stands in stark contrast to its soft, creamy white belly. Its wings sweep upward with breathtaking force, fully extended. The outline of each feather is clearly visible in the air, and the dust and wall debris that rise up create a fleeting halo of light around it. The color saturation of the entire picture is extremely high, and the light makes the warm tones of the wall particularly mellow, forming a distinct and harmonious visual impact with the cool luster of the swallow's feathers, creating a strong atmosphere of vitality bursting out in chaos and capturing eternity in the fleeting moment.", + "seed": 19479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19479_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010148442396894098, + "vlm_logprob": -0.010148442396894098, + "vlm_yes_no_margin": 5.925, + "vlm_token_credit": -0.010148442396894098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001353539526462555 + ], + "opposite_logprob_tokens": [ + -7.751353740692139 + ], + "logprob_reward": -0.001353539526462555, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.001353539526462555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149488165974617, + 0.08751283586025238, + 0.05339512228965759, + 0.09062724560499191, + 0.09493479132652283, + 0.10811938345432281, + 0.11244522780179977, + 0.12578992545604706, + 0.12940582633018494, + 0.04828151315450668 + ], + "credit_map": [ + 0.01236874982714653, + 0.022123269736766815, + 0.013667602092027664, + 0.015109354630112648, + 0.025199733674526215, + 0.028609609231352806, + 0.018248291686177254, + 0.014161556027829647, + 0.01159969437867403, + 0.014336025342345238, + 0.006857161410152912, + 0.013017126359045506, + 0.01247350312769413, + 0.01305670477449894, + 0.0072007859125733376, + 0.008971834555268288, + 0.00525256572291255, + 0.008369997143745422, + 0.006538621615618467, + 0.004523029550909996, + 0.007418086752295494, + 0.008234546519815922, + 0.005494715180248022, + 0.00756355794146657, + 0.009161589667201042, + 0.02340269647538662, + 0.008011574856936932, + 0.00433250842615962, + 0.02435225434601307, + 0.010357935912907124, + 0.004311093594878912, + 0.006697589531540871, + 0.01259161438792944, + 0.01312419306486845, + 0.011565952561795712, + 0.010530115105211735, + 0.013390602543950081, + 0.015426676720380783, + 0.010429730638861656, + 0.007875902578234673, + 0.01870601251721382, + 0.01338899228721857, + 0.027780191972851753, + 0.007090717554092407, + 0.007544097024947405, + 0.014655066654086113, + 0.00911568384617567, + 0.009838628582656384, + 0.027491167187690735, + 0.012227639555931091, + 0.04001598432660103, + 0.00893067941069603, + 0.005509041249752045, + 0.005889662075787783, + 0.008049330674111843, + 0.0043317354284226894, + 0.02351566217839718, + 0.01525829080492258, + 0.035812489688396454, + 0.009987957775592804, + 0.010745102539658546, + 0.013178330846130848, + 0.012090628035366535, + 0.00520147243514657, + 0.02779025211930275, + 0.02641388401389122, + 0.022329920902848244, + 0.012383658438920975, + 0.011991055682301521, + 0.010468387976288795, + 0.0112900510430336, + 0.006738615222275257, + 0.0049153659492731094, + 0.007102217059582472, + 0.0058756424114108086, + 0.005327719263732433, + 0.008154475130140781, + 0.005666477605700493, + 0.007419423200190067, + 0.0038201920688152313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.001353539526462555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819374669343233 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819374669343233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819374669343233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25122785568237305, + 0.019369205459952354, + 0.17045630514621735, + 0.3988669812679291, + 0.018954643979668617, + 0.018246224150061607, + 0.049670375883579254, + 0.02625771053135395, + 0.03037731721997261, + 0.016573432832956314 + ], + "credit_map": [ + 0.018884364515542984, + 0.04006321355700493, + 0.018240101635456085, + 0.017139462754130363, + 0.04720417037606239, + 0.055370889604091644, + 0.03173283487558365, + 0.02259279228746891, + 0.002739767776802182, + 0.0037755996454507113, + 0.0021092710085213184, + 0.0021046199835836887, + 0.0026411761064082384, + 0.0027015293017029762, + 0.0017117790412157774, + 0.0015854599187150598, + 0.01453374046832323, + 0.02366345003247261, + 0.020532887428998947, + 0.02460743673145771, + 0.026332782581448555, + 0.029438437893986702, + 0.010766884312033653, + 0.02058066986501217, + 0.013734322041273117, + 0.16895316541194916, + 0.004422825761139393, + 0.00417229114100337, + 0.18120282888412476, + 0.016633547842502594, + 0.0045508029870688915, + 0.005197151098400354, + 0.0022773572709411383, + 0.0025367180351167917, + 0.0029501819517463446, + 0.002009462332352996, + 0.002386382780969143, + 0.002556453924626112, + 0.0024211183190345764, + 0.0018169672694057226, + 0.0025878101587295532, + 0.002020796062424779, + 0.0040918271988630295, + 0.0020657603163272142, + 0.0017226155614480376, + 0.002418245654553175, + 0.0014886903809383512, + 0.001850475324317813, + 0.010564139112830162, + 0.005824225954711437, + 0.01639791578054428, + 0.0035813727881759405, + 0.0029940831009298563, + 0.004004483576864004, + 0.0035872245207428932, + 0.002716922899708152, + 0.004469173029065132, + 0.0051204306073486805, + 0.006465633399784565, + 0.0024100332520902157, + 0.0025327319744974375, + 0.0017838020576164126, + 0.002416962292045355, + 0.001058940659277141, + 0.0036519006825983524, + 0.00524409394711256, + 0.006501401774585247, + 0.0038796975277364254, + 0.0032810186967253685, + 0.0032488135620951653, + 0.0027918554842472076, + 0.0017785339150577784, + 0.0015823548892512918, + 0.0019389567896723747, + 0.0028265132568776608, + 0.0025286723393946886, + 0.0018273696769028902, + 0.002114212140440941, + 0.00237101293168962, + 0.001384338946081698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.002819374669343233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010409895330667496 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010409895330667496, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010409895330667496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03514786437153816, + 0.034175168722867966, + 0.036340489983558655, + 0.5426020622253418, + 0.032709088176488876, + 0.02113039791584015, + 0.10935978591442108, + 0.0970212072134018, + 0.06919148564338684, + 0.022322535514831543 + ], + "credit_map": [ + 0.0035582189448177814, + 0.00578497676178813, + 0.0036183143965899944, + 0.004280645400285721, + 0.0048184809274971485, + 0.005778943188488483, + 0.0038110592868179083, + 0.0034972254652529955, + 0.006147395353764296, + 0.010292913764715195, + 0.0022962274961173534, + 0.003351981518790126, + 0.004194663837552071, + 0.0047129616141319275, + 0.001392891863361001, + 0.001786131993867457, + 0.00313576846383512, + 0.00460735522210598, + 0.0052622533403337, + 0.006571603007614613, + 0.005405302159488201, + 0.004221176728606224, + 0.0028455944266170263, + 0.004291432909667492, + 0.02043195441365242, + 0.21945121884346008, + 0.011418621055781841, + 0.010683194734156132, + 0.23839445412158966, + 0.024040918797254562, + 0.007749488111585379, + 0.010432160459458828, + 0.004307052586227655, + 0.0039392998442053795, + 0.004727167077362537, + 0.0024210873525589705, + 0.003382462775334716, + 0.008227907121181488, + 0.0029613208025693893, + 0.0027427871245890856, + 0.0035869241692125797, + 0.002499896101653576, + 0.003792419331148267, + 0.0013205221621319652, + 0.0014558978145942092, + 0.004345147870481014, + 0.0013401900650933385, + 0.0027893981896340847, + 0.025441383942961693, + 0.009655284695327282, + 0.04102544113993645, + 0.009254495613276958, + 0.005613218527287245, + 0.006715533789247274, + 0.006803752388805151, + 0.0048506720922887325, + 0.013560609892010689, + 0.012639924883842468, + 0.02155882492661476, + 0.007120963651686907, + 0.009958830662071705, + 0.014107566326856613, + 0.0144644221290946, + 0.0036100544966757298, + 0.012428443878889084, + 0.013851543888449669, + 0.009267683140933514, + 0.005285670980811119, + 0.010661641135811806, + 0.0073294141329824924, + 0.006726376712322235, + 0.003640709677711129, + 0.002665039151906967, + 0.003203559899702668, + 0.005797171033918858, + 0.0025573670864105225, + 0.002637474797666073, + 0.002192841377109289, + 0.0017575668171048164, + 0.0015115122077986598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.0010409895330667496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024839758407324553 + ], + "opposite_logprob_tokens": [ + -6.00248384475708 + ], + "logprob_reward": -0.0024839758407324553, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024839758407324553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06709159910678864, + 0.023242631927132607, + 0.021241771057248116, + 0.5410995483398438, + 0.02957475557923317, + 0.12147272378206253, + 0.08228571712970734, + 0.06639163941144943, + 0.021382974460721016, + 0.026216646656394005 + ], + "credit_map": [ + 0.0060585541650652885, + 0.011758163571357727, + 0.0050453138537704945, + 0.006069639232009649, + 0.009878965094685555, + 0.013029513880610466, + 0.008616735227406025, + 0.00663471594452858, + 0.003718275809660554, + 0.004788871388882399, + 0.002055021934211254, + 0.00214643613435328, + 0.003062254749238491, + 0.003184353234246373, + 0.0026498830411583185, + 0.0016375380801036954, + 0.002584518399089575, + 0.004747196100652218, + 0.001726440154016018, + 0.001345247495919466, + 0.0033922470174729824, + 0.0033869168255478144, + 0.001951164216734469, + 0.0021080414298921824, + 0.01807289570569992, + 0.22940678894519806, + 0.007262101396918297, + 0.005995121784508228, + 0.2449086606502533, + 0.021891990676522255, + 0.006227951962500811, + 0.007334029767662287, + 0.0034795731771737337, + 0.004556918982416391, + 0.0038820013869553804, + 0.001991220749914646, + 0.003934734966605902, + 0.007024110294878483, + 0.0025313040241599083, + 0.0021748915314674377, + 0.020010679960250854, + 0.014008935540914536, + 0.02751036174595356, + 0.011264086700975895, + 0.009198036044836044, + 0.01756623201072216, + 0.011422349140048027, + 0.010492038913071156, + 0.020702030509710312, + 0.0068666269071400166, + 0.03338746353983879, + 0.007573516108095646, + 0.0030691896099597216, + 0.0037173605524003506, + 0.004754731431603432, + 0.0022147921845316887, + 0.013848694041371346, + 0.009344483725726604, + 0.020391207188367844, + 0.005423011723905802, + 0.004399631638079882, + 0.005426537711173296, + 0.005566348787397146, + 0.0019917297177016735, + 0.002907385118305683, + 0.0038590270560234785, + 0.0033710277639329433, + 0.0023766953963786364, + 0.0033558178693056107, + 0.0019046756206080317, + 0.002289725234732032, + 0.0013186200521886349, + 0.0027344906702637672, + 0.003200915176421404, + 0.005909357685595751, + 0.0030811852775514126, + 0.0029664793983101845, + 0.002857733517885208, + 0.0034334210213273764, + 0.0020330646075308323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.0024839758407324553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043044332414865494 + ], + "opposite_logprob_tokens": [ + -3.168044328689575 + ], + "logprob_reward": -0.043044332414865494, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043044332414865494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13427308201789856, + 0.08425939828157425, + 0.015895944088697433, + 0.12637512385845184, + 0.04682080075144768, + 0.1571965515613556, + 0.21122997999191284, + 0.08651507645845413, + 0.0831804946064949, + 0.05425359308719635 + ], + "credit_map": [ + 0.01049390435218811, + 0.014714241959154606, + 0.009356540627777576, + 0.009238566271960735, + 0.03311214596033096, + 0.037462811917066574, + 0.007602139841765165, + 0.01229273620992899, + 0.018431326374411583, + 0.01916487328708172, + 0.006269725039601326, + 0.006006103940308094, + 0.010798624716699123, + 0.012953538447618484, + 0.0057275621220469475, + 0.004907643888145685, + 0.0017366381362080574, + 0.002445951336994767, + 0.0017215015832334757, + 0.0019941942300647497, + 0.0025441390462219715, + 0.002169466344639659, + 0.0015300747472792864, + 0.0017539780819788575, + 0.005350559949874878, + 0.04712802544236183, + 0.004041676875203848, + 0.003502455074340105, + 0.052819330245256424, + 0.008037104271352291, + 0.0023355153389275074, + 0.0031604734249413013, + 0.00469932472333312, + 0.005698313470929861, + 0.0100270239636302, + 0.004609052557498217, + 0.004582913592457771, + 0.009371926076710224, + 0.0038135002832859755, + 0.004018744919449091, + 0.02452690899372101, + 0.0179748572409153, + 0.03903721645474434, + 0.01041688583791256, + 0.009443407878279686, + 0.027347782626748085, + 0.00925994198769331, + 0.01918955147266388, + 0.04961878061294556, + 0.026103047654032707, + 0.07556578516960144, + 0.015575332567095757, + 0.01170208677649498, + 0.010867101140320301, + 0.014989527873694897, + 0.006808301899582148, + 0.02044956013560295, + 0.00857431162148714, + 0.029761996120214462, + 0.0069900937378406525, + 0.005118285771459341, + 0.005400525406002998, + 0.00729769142344594, + 0.0029226175975054502, + 0.013302790932357311, + 0.015123951248824596, + 0.01718372292816639, + 0.00857997965067625, + 0.009325431659817696, + 0.007017051801085472, + 0.007906540296971798, + 0.004741025622934103, + 0.006690169684588909, + 0.006334126926958561, + 0.01102594193071127, + 0.005147308576852083, + 0.008214755915105343, + 0.0070914290845394135, + 0.0059169139713048935, + 0.0038329458329826593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.043044332414865494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Steve McCurry's signature rich color style, a slow-motion medium shot, captured with high-speed cinematography, precisely freezes a moment brimming with dramatic tension. A warm, intense afternoon light slants down, illuminating a weathered, mottled wall. The surface is a rich interplay of deep ochre yellow and faded turquoise blue. The paint, eroded by age, has flaked off, revealing the rough plaster beneath. Before this storied wall, a graceful swallow, startled by an unknown source, suddenly takes flight, its body tensed in the air. Caught in extreme slow motion, its back, shimmering with metallic blue-green feathers, stands in stark contrast to its soft, creamy white belly. Its wings sweep upward with breathtaking force, fully extended. The outline of each feather is clearly visible in the air, and the dust and wall debris that rise up create a fleeting halo of light around it. The color saturation of the entire picture is extremely high, and the light makes the warm tones of the wall particularly mellow, forming a distinct and harmonious visual impact with the cool luster of the swallow's feathers, creating a strong atmosphere of vitality bursting out in chaos and capturing eternity in the fleeting moment.", + "seed": 19480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_10/index_9265_seed_19480_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.007423352543264627, + "vlm_logprob": -0.007423352543264627, + "vlm_yes_no_margin": 5.8, + "vlm_token_credit": -0.007423352543264627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028307863976806402 + ], + "opposite_logprob_tokens": [ + -7.377830982208252 + ], + "logprob_reward": -0.0028307863976806402, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0028307863976806402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515699177980423, + 0.11033911257982254, + 0.06584775447845459, + 0.04912229999899864, + 0.08727666735649109, + 0.09301045536994934, + 0.19318030774593353, + 0.07813660055398941, + 0.0772947147488594, + 0.09422220289707184 + ], + "credit_map": [ + 0.02024410292506218, + 0.015613722614943981, + 0.019632438197731972, + 0.020261699333786964, + 0.02349027991294861, + 0.030187195166945457, + 0.011505800299346447, + 0.010634683072566986, + 0.008503327146172523, + 0.012843860313296318, + 0.01194704044610262, + 0.013843574561178684, + 0.01785539649426937, + 0.019062554463744164, + 0.012814095243811607, + 0.013469258323311806, + 0.00651611527428031, + 0.0065900664776563644, + 0.013266508467495441, + 0.00859962310642004, + 0.008590992540121078, + 0.0069505698047578335, + 0.007742845453321934, + 0.007591025438159704, + 0.004083399660885334, + 0.006424793507903814, + 0.006094612646847963, + 0.00879521295428276, + 0.004345641005784273, + 0.00886434968560934, + 0.004437156021595001, + 0.006077134981751442, + 0.005325788166373968, + 0.009715818800032139, + 0.016822783276438713, + 0.02227281965315342, + 0.00766251003369689, + 0.010572969913482666, + 0.006821153219789267, + 0.008082825690507889, + 0.008390095084905624, + 0.014832776971161366, + 0.009171291254460812, + 0.014674230478703976, + 0.012261452153325081, + 0.0085151931270957, + 0.011431093327701092, + 0.013734322041273117, + 0.03129637986421585, + 0.021094851195812225, + 0.022387556731700897, + 0.0193229578435421, + 0.021719960495829582, + 0.037340033799409866, + 0.01845189929008484, + 0.02156667225062847, + 0.007936113514006138, + 0.009181017056107521, + 0.010717872530221939, + 0.007213361095637083, + 0.01413747202605009, + 0.010512693785130978, + 0.012692631222307682, + 0.005745438858866692, + 0.011489287950098515, + 0.009255139157176018, + 0.007636863272637129, + 0.007610710803419352, + 0.013805617578327656, + 0.009997564367949963, + 0.0096249645575881, + 0.007874572649598122, + 0.010761317797005177, + 0.010547070764005184, + 0.011761746369302273, + 0.016576794907450676, + 0.01331001240760088, + 0.008910277858376503, + 0.014381169341504574, + 0.007973806001245975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0028307863976806402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.549322247505188, + 0.039606980979442596, + 0.05279381200671196, + 0.059683993458747864, + 0.07324191182851791, + 0.0700945034623146, + 0.022055448964238167, + 0.02623934857547283, + 0.06998591125011444, + 0.036975860595703125 + ], + "credit_map": [ + 0.03306616097688675, + 0.021039610728621483, + 0.018132077530026436, + 0.03769543394446373, + 0.1775272786617279, + 0.21251840889453888, + 0.017146052792668343, + 0.03219719976186752, + 0.0036519363056868315, + 0.004632133059203625, + 0.004512563347816467, + 0.004406544379889965, + 0.004601779859513044, + 0.005113271530717611, + 0.008472966961562634, + 0.004215779714286327, + 0.005988733377307653, + 0.006563471630215645, + 0.009389619342982769, + 0.008288444019854069, + 0.006210422609001398, + 0.004782044794410467, + 0.006373092532157898, + 0.005197979509830475, + 0.003595517948269844, + 0.003937337081879377, + 0.005527128465473652, + 0.02075899951159954, + 0.0035882489755749702, + 0.008826259523630142, + 0.0038861085195094347, + 0.009564382024109364, + 0.00454874848946929, + 0.008549986407160759, + 0.01128422748297453, + 0.018688609823584557, + 0.006123509723693132, + 0.009944241493940353, + 0.007208743132650852, + 0.006893839221447706, + 0.009116997942328453, + 0.01079082302749157, + 0.007322842720896006, + 0.01041487418115139, + 0.00922217033803463, + 0.0059255994856357574, + 0.008105363696813583, + 0.009195826947689056, + 0.0025365077890455723, + 0.002274635247886181, + 0.0032401110511273146, + 0.002072257688269019, + 0.0029945934657007456, + 0.003380535403266549, + 0.003044841578230262, + 0.002511963713914156, + 0.004116587806493044, + 0.0036163756158202887, + 0.0035138193052262068, + 0.004182070959359407, + 0.002653895178809762, + 0.0026260046288371086, + 0.003171851160004735, + 0.0023587418254464865, + 0.014984154142439365, + 0.007906650193035603, + 0.006521227303892374, + 0.007917145267128944, + 0.006325239781290293, + 0.006831699516624212, + 0.008335151709616184, + 0.01116463728249073, + 0.005126742646098137, + 0.003769038012251258, + 0.00714927027001977, + 0.004717908799648285, + 0.003927124198526144, + 0.002515810076147318, + 0.007014446426182985, + 0.00275551644153893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00151305110193789 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.00151305110193789, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00151305110193789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14764250814914703, + 0.06505922973155975, + 0.0636548325419426, + 0.2392863780260086, + 0.1683977097272873, + 0.038778334856033325, + 0.07753463834524155, + 0.05781581252813339, + 0.07815144956111908, + 0.06367912143468857 + ], + "credit_map": [ + 0.010226148180663586, + 0.00789417140185833, + 0.01102360337972641, + 0.01658809185028076, + 0.03840629383921623, + 0.03849266842007637, + 0.009820185601711273, + 0.015191343612968922, + 0.005045333877205849, + 0.00751855643466115, + 0.007049807347357273, + 0.009554331190884113, + 0.008989309892058372, + 0.007631551939994097, + 0.011376611888408661, + 0.007893729954957962, + 0.006614895537495613, + 0.005759556777775288, + 0.009221473708748817, + 0.009825514629483223, + 0.0068559544160962105, + 0.00815579853951931, + 0.010169316083192825, + 0.007052322383970022, + 0.014150538481771946, + 0.015565750189125538, + 0.024097144603729248, + 0.07356371730566025, + 0.019546642899513245, + 0.05289235711097717, + 0.019329609349370003, + 0.020140619948506355, + 0.010217273607850075, + 0.024346472695469856, + 0.02082735486328602, + 0.021843787282705307, + 0.025629475712776184, + 0.0284279678016901, + 0.019593778997659683, + 0.017511600628495216, + 0.002594318240880966, + 0.004868500865995884, + 0.0032736477442085743, + 0.006698446813970804, + 0.005668857134878635, + 0.006001647096127272, + 0.004869162570685148, + 0.004803752060979605, + 0.011462902650237083, + 0.01012086309492588, + 0.00980308372527361, + 0.009763859212398529, + 0.007078579626977444, + 0.011092445813119411, + 0.009974444285035133, + 0.008238463662564754, + 0.006770147010684013, + 0.007322236895561218, + 0.006413616240024567, + 0.00846923515200615, + 0.008241045288741589, + 0.007269039284437895, + 0.007890235632658005, + 0.0054402598179876804, + 0.007196473889052868, + 0.008711411617696285, + 0.007391298655420542, + 0.010710627771914005, + 0.015277962200343609, + 0.009034163318574429, + 0.012822184711694717, + 0.007007323671132326, + 0.009326848201453686, + 0.007948385551571846, + 0.01308277528733015, + 0.005755096673965454, + 0.008660539984703064, + 0.004385202657431364, + 0.009320211596786976, + 0.005200056824833155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.00151305110193789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003612422849982977 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.003612422849982977, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003612422849982977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1907360851764679, + 0.04463398456573486, + 0.06700783222913742, + 0.08802404999732971, + 0.06668328493833542, + 0.14669886231422424, + 0.15134480595588684, + 0.0574127621948719, + 0.11121875792741776, + 0.07623957097530365 + ], + "credit_map": [ + 0.01530687790364027, + 0.01113280188292265, + 0.011994936503469944, + 0.01715569756925106, + 0.048569049686193466, + 0.059795256704092026, + 0.010267574340105057, + 0.01651388593018055, + 0.006162402685731649, + 0.0037662091199308634, + 0.005646736826747656, + 0.0051736971363425255, + 0.00498233363032341, + 0.005353402812033892, + 0.008809800259768963, + 0.0047394027933478355, + 0.0069100684486329556, + 0.008722017519176006, + 0.010426415130496025, + 0.009368578903377056, + 0.009295069612562656, + 0.007804749067872763, + 0.008086124435067177, + 0.006394808180630207, + 0.006775734480470419, + 0.011538459919393063, + 0.009957571513950825, + 0.015833944082260132, + 0.008449855260550976, + 0.011896601878106594, + 0.00807797722518444, + 0.0154939079657197, + 0.005954022519290447, + 0.009428870864212513, + 0.009579871781170368, + 0.00907245185226202, + 0.008915564976632595, + 0.008843885734677315, + 0.007110386621206999, + 0.0077782245352864265, + 0.011893193237483501, + 0.021144693717360497, + 0.010887524113059044, + 0.026355767622590065, + 0.023430531844496727, + 0.011854014359414577, + 0.02039356902241707, + 0.020739564672112465, + 0.018972918391227722, + 0.020396992564201355, + 0.021648606285452843, + 0.009872891940176487, + 0.017641140148043633, + 0.03281586989760399, + 0.013346796855330467, + 0.016649603843688965, + 0.010704204440116882, + 0.007894079200923443, + 0.007068766746670008, + 0.007146703079342842, + 0.00767335994169116, + 0.004194667097181082, + 0.007364718709141016, + 0.005366265308111906, + 0.01338205672800541, + 0.013120514340698719, + 0.009150851517915726, + 0.010980416089296341, + 0.016728797927498817, + 0.018108751624822617, + 0.01294570043683052, + 0.01680167019367218, + 0.010992537252604961, + 0.007599478587508202, + 0.016381599009037018, + 0.010648898780345917, + 0.006880517583340406, + 0.004748363047838211, + 0.013260272331535816, + 0.005727906245738268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.003612422849982977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set before a weathered wall with deep ochre yellow and faded turquoise blue paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wall surface eroded with flaked paint revealing rough plaster beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a graceful swallow visible with metallic blue-green back feathers and a creamy white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swallow captured in extreme slow motion with wings fully extended and sweeping upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color saturation extremely high with warm wall tones and cool feather luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339344680309296 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339344680309296, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339344680309296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23713290691375732, + 0.062101587653160095, + 0.02262698858976364, + 0.10781027376651764, + 0.07715292274951935, + 0.041111670434474945, + 0.083034947514534, + 0.19518935680389404, + 0.12312731146812439, + 0.05071205645799637 + ], + "credit_map": [ + 0.014611836522817612, + 0.011286657303571701, + 0.012700799852609634, + 0.01727449893951416, + 0.07035607844591141, + 0.08796937763690948, + 0.010312385857105255, + 0.01262127235531807, + 0.004530912730842829, + 0.007720437366515398, + 0.007834778167307377, + 0.008432463742792606, + 0.008427537977695465, + 0.007896251045167446, + 0.011327170766890049, + 0.005932039115577936, + 0.0020846372935920954, + 0.002996762515977025, + 0.003985571209341288, + 0.002872886136174202, + 0.0027053856756538153, + 0.0024463660083711147, + 0.002861724002286792, + 0.0026736578438431025, + 0.007030829321593046, + 0.008007164113223553, + 0.009801867417991161, + 0.03699548915028572, + 0.00431510666385293, + 0.022597478702664375, + 0.006216010078787804, + 0.012846327386796474, + 0.0055289845913648605, + 0.00950847938656807, + 0.012255151756107807, + 0.015989137813448906, + 0.007450175937265158, + 0.013585373759269714, + 0.006972689647227526, + 0.005862930323928595, + 0.006028271745890379, + 0.004382368177175522, + 0.004246204160153866, + 0.0042976681143045425, + 0.0060959551483392715, + 0.003858366282656789, + 0.005141598172485828, + 0.00706124072894454, + 0.00988298561424017, + 0.008221092633903027, + 0.011185816489160061, + 0.010323083028197289, + 0.01217262726277113, + 0.014446702785789967, + 0.0071973418816924095, + 0.009605295024812222, + 0.030347177758812904, + 0.022547205910086632, + 0.025431308895349503, + 0.036337558180093765, + 0.021819980815052986, + 0.01530328206717968, + 0.025988871231675148, + 0.017413966357707977, + 0.017642538994550705, + 0.010332346893846989, + 0.013659235090017319, + 0.026923613622784615, + 0.008368507958948612, + 0.010078571736812592, + 0.01811687834560871, + 0.018005620688199997, + 0.009667320176959038, + 0.006179049611091614, + 0.012034681625664234, + 0.005886314436793327, + 0.004767840262502432, + 0.002780779730528593, + 0.006242831703275442, + 0.0031532375141978264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.026339344680309296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_tensor([2093, 2093])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_tensor([2093, 2093])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bc56a212003a6f11ececc47308a73ee8e611de1a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_tensor([2093, 2093])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-shot shot, taken from a level angle, draws the viewer's gaze deep into the depths of a dimly lit, ancient inn. At the center of the frame, a lone swordsman sits at a greasy, square wooden table. He wears a dark blue, coarse cloth gown, slightly frayed at the collar and cuffs. His long, jet-black hair is casually tied back with a simple cloth sash, a few strands falling across his sharply defined features. A flickering oil lamp on the table provides the only light source. The warm, dim glow casts deep shadows on his face, highlighting his tightly pursed lips and eagle-like gaze. He leans forward slightly, his gaze fixed intently on the heavy wooden door not far away, as if awaiting the arrival of a fate. His right hand, with distinct joints and veins visible, grips the hilt of an ancient longsword at his waist with a gesture imbued with immense power. The hilt is made of dark wood and wrapped in aged leather. The main colors of the entire scene are dark brown and dark red, illuminated by that jumping orange light. The air is filled with a silent, focused and tense atmosphere that is about to explode, like the last calm before the storm.", + "seed": 6485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_2093_seed_6485_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015002926345914602, + "vlm_logprob": -0.015002926345914602, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.015002926345914602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032008010894060135 + ], + "opposite_logprob_tokens": [ + -3.532007932662964 + ], + "logprob_reward": -0.032008010894060135, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.032008010894060135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3154008388519287, + 0.11080406606197357, + 0.0768476128578186, + 0.05970843508839607, + 0.104477159678936, + 0.09398447722196579, + 0.04324587807059288, + 0.032771024852991104, + 0.04597604647278786, + 0.1167844608426094 + ], + "credit_map": [ + 0.029174871742725372, + 0.052629504352808, + 0.020641600713133812, + 0.032055944204330444, + 0.03552355617284775, + 0.0628703162074089, + 0.0389898456633091, + 0.04351521283388138, + 0.01482468843460083, + 0.01465944666415453, + 0.01532295998185873, + 0.010428130626678467, + 0.009398676455020905, + 0.022161632776260376, + 0.012212986126542091, + 0.011795535683631897, + 0.008600970730185509, + 0.011899826116859913, + 0.006222677882760763, + 0.010354344733059406, + 0.009846480563282967, + 0.012363540939986706, + 0.009204206988215446, + 0.008355566300451756, + 0.008905407041311264, + 0.00795292854309082, + 0.005759530700743198, + 0.007721274625509977, + 0.006234666332602501, + 0.008962639607489109, + 0.006090345326811075, + 0.008081641979515553, + 0.01522843074053526, + 0.01965983584523201, + 0.008765996433794498, + 0.012678244151175022, + 0.012158398516476154, + 0.015860948711633682, + 0.010012934915721416, + 0.010112368501722813, + 0.011919943615794182, + 0.016545401886105537, + 0.008525338023900986, + 0.009933562949299812, + 0.010919434949755669, + 0.016186252236366272, + 0.010247678495943546, + 0.009706868790090084, + 0.005320644937455654, + 0.005730618257075548, + 0.004468850325793028, + 0.005792051553726196, + 0.004958219360560179, + 0.007421053480356932, + 0.004483184311538935, + 0.005071256775408983, + 0.004166190046817064, + 0.0037785922177135944, + 0.003465821500867605, + 0.005142327398061752, + 0.004071495961397886, + 0.004570756573230028, + 0.0037312707863748074, + 0.003844571067020297, + 0.004237826447933912, + 0.004291638266295195, + 0.002368737244978547, + 0.0059991925954818726, + 0.006185528822243214, + 0.006814336404204369, + 0.010020194575190544, + 0.006058593280613422, + 0.014249587431550026, + 0.0187902320176363, + 0.010152677074074745, + 0.014355470426380634, + 0.012960479594767094, + 0.019345683977007866, + 0.015618362464010715, + 0.011311963200569153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.032008010894060135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205054379999638 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205054379999638, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205054379999638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2655593156814575, + 0.0928822010755539, + 0.08231616020202637, + 0.07306485623121262, + 0.1358499377965927, + 0.048815738409757614, + 0.050708405673503876, + 0.07829319685697556, + 0.09338998049497604, + 0.07912015914916992 + ], + "credit_map": [ + 0.02456694468855858, + 0.03703780099749565, + 0.018792742863297462, + 0.027400851249694824, + 0.05102061480283737, + 0.04216695576906204, + 0.03673200681805611, + 0.027841417118906975, + 0.010509183630347252, + 0.010106310248374939, + 0.006075575947761536, + 0.011956445872783661, + 0.014525257982313633, + 0.013625327497720718, + 0.013146528042852879, + 0.01293757651001215, + 0.008141481317579746, + 0.006941971834748983, + 0.00583452545106411, + 0.009887185879051685, + 0.012898514047265053, + 0.014098119921982288, + 0.011735511012375355, + 0.012778845615684986, + 0.0053422036580741405, + 0.00834636576473713, + 0.0048908572643995285, + 0.013132734224200249, + 0.01140877977013588, + 0.008981999009847641, + 0.01227024756371975, + 0.008691668510437012, + 0.010188105516135693, + 0.011227846145629883, + 0.00526195764541626, + 0.013825382106006145, + 0.02742832526564598, + 0.019747024402022362, + 0.02538691833615303, + 0.022784389555454254, + 0.006225832737982273, + 0.006341367028653622, + 0.004705994389951229, + 0.00585503363981843, + 0.0051788631826639175, + 0.008827549405395985, + 0.005488706286996603, + 0.006192395463585854, + 0.005687105003744364, + 0.006186651065945625, + 0.004503482952713966, + 0.005662126466631889, + 0.006548506673425436, + 0.008628908544778824, + 0.0066751958802342415, + 0.006816430017352104, + 0.007662083953619003, + 0.0075740693137049675, + 0.004516413901001215, + 0.010083266533911228, + 0.009468505159020424, + 0.01016328763216734, + 0.013147340156137943, + 0.01567823253571987, + 0.011152035556733608, + 0.012285309843719006, + 0.011006627231836319, + 0.01052496675401926, + 0.012689048424363136, + 0.014602151699364185, + 0.01021168939769268, + 0.01091816183179617, + 0.006952115800231695, + 0.007847866043448448, + 0.005958291236311197, + 0.009924746118485928, + 0.012874459847807884, + 0.011503308080136776, + 0.010680891573429108, + 0.013378488831222057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.014205054379999638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31494706869125366, + 0.12400629371404648, + 0.06500739604234695, + 0.049750182777643204, + 0.05979849770665169, + 0.10118450969457626, + 0.07709517329931259, + 0.08705427497625351, + 0.06170656532049179, + 0.059449996799230576 + ], + "credit_map": [ + 0.026645159348845482, + 0.03705839812755585, + 0.02726477012038231, + 0.04219265654683113, + 0.05470799654722214, + 0.041631121188402176, + 0.04067055135965347, + 0.04477643966674805, + 0.01063720602542162, + 0.016464892774820328, + 0.0075493426993489265, + 0.020033441483974457, + 0.021393172442913055, + 0.015993189066648483, + 0.01755591109395027, + 0.014379139058291912, + 0.005171474535018206, + 0.008300801739096642, + 0.003549676388502121, + 0.010998326353728771, + 0.012610627338290215, + 0.008662721142172813, + 0.008458017371594906, + 0.007255756761878729, + 0.004336146637797356, + 0.0056279366835951805, + 0.0024721678346395493, + 0.00745406886562705, + 0.007979120127856731, + 0.007610560394823551, + 0.007955250330269337, + 0.0063149346970021725, + 0.005075704772025347, + 0.00856352411210537, + 0.0031199308577924967, + 0.0070226313546299934, + 0.00846797414124012, + 0.008115705102682114, + 0.007576275616884232, + 0.01185675524175167, + 0.010569280944764614, + 0.012283393181860447, + 0.006043981295078993, + 0.010728888213634491, + 0.016463369131088257, + 0.013347827829420567, + 0.016556113958358765, + 0.015191661193966866, + 0.006868686061352491, + 0.007179780397564173, + 0.004087407141923904, + 0.009833140298724174, + 0.012203645892441273, + 0.010987919755280018, + 0.013122878968715668, + 0.012811718508601189, + 0.006219110451638699, + 0.007424972951412201, + 0.00502448994666338, + 0.011883183382451534, + 0.013482530601322651, + 0.012066260911524296, + 0.015688523650169373, + 0.015265213325619698, + 0.005670327693223953, + 0.008633406832814217, + 0.003986178897321224, + 0.009292270056903362, + 0.008798172697424889, + 0.008868948556482792, + 0.007930695079267025, + 0.008526571094989777, + 0.005962237250059843, + 0.007911466993391514, + 0.0051241242326796055, + 0.007730556186288595, + 0.009322833269834518, + 0.008152931928634644, + 0.0077618081122636795, + 0.007484044414013624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095263313502073 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095263313502073, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095263313502073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3279009461402893, + 0.1129114031791687, + 0.08282774686813354, + 0.08389946073293686, + 0.07083546370267868, + 0.05321063846349716, + 0.10464461147785187, + 0.05227919667959213, + 0.08592431247234344, + 0.025566261261701584 + ], + "credit_map": [ + 0.03881193324923515, + 0.029694512486457825, + 0.031268805265426636, + 0.04317958280444145, + 0.030411234125494957, + 0.0547938197851181, + 0.048908330500125885, + 0.050832729786634445, + 0.012804958038032055, + 0.014071539975702763, + 0.008945109322667122, + 0.014500417746603489, + 0.012542263604700565, + 0.016479268670082092, + 0.01715574413537979, + 0.01641208678483963, + 0.011173374019563198, + 0.010009469464421272, + 0.0059938314370810986, + 0.011303305625915527, + 0.010160819627344608, + 0.012032895348966122, + 0.011768043041229248, + 0.010385998524725437, + 0.010320883244276047, + 0.009456074796617031, + 0.00813811644911766, + 0.011562717147171497, + 0.011433910578489304, + 0.012144730426371098, + 0.011410439386963844, + 0.009432580322027206, + 0.005504019092768431, + 0.00628344202414155, + 0.0037439700681716204, + 0.00967920571565628, + 0.009355941787362099, + 0.010323015041649342, + 0.008816211484372616, + 0.017129648476839066, + 0.005412486847490072, + 0.0068763671442866325, + 0.004913774784654379, + 0.0055670118890702724, + 0.006339140702039003, + 0.007852035574615002, + 0.005499251186847687, + 0.01075056754052639, + 0.013202096335589886, + 0.016518792137503624, + 0.011950881220400333, + 0.011675500310957432, + 0.00991914514452219, + 0.017601432278752327, + 0.011088516563177109, + 0.01268824003636837, + 0.005856991745531559, + 0.005638939328491688, + 0.004153743851929903, + 0.007144897244870663, + 0.005787618458271027, + 0.007715035229921341, + 0.008316539227962494, + 0.007665424142032862, + 0.010730776004493237, + 0.009152601473033428, + 0.007005487568676472, + 0.011201239190995693, + 0.007658038288354874, + 0.012625504285097122, + 0.015007847920060158, + 0.01254281122237444, + 0.00278542865999043, + 0.0026071807369589806, + 0.002414596965536475, + 0.004052645526826382, + 0.004085981752723455, + 0.003547368571162224, + 0.0033826963044703007, + 0.0026903594844043255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004095263313502073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606672629714012 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.01606672629714012, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606672629714012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3125889003276825, + 0.14170482754707336, + 0.07000900059938431, + 0.0809403732419014, + 0.069632388651371, + 0.06392960995435715, + 0.06720956414937973, + 0.05029669031500816, + 0.1057903990149498, + 0.037898238748311996 + ], + "credit_map": [ + 0.029137417674064636, + 0.03408835828304291, + 0.029769960790872574, + 0.057710833847522736, + 0.046486932784318924, + 0.03693617135286331, + 0.03849411755800247, + 0.03996511548757553, + 0.013228562660515308, + 0.015594357624650002, + 0.007805669214576483, + 0.026253076270222664, + 0.018281495198607445, + 0.018853364512324333, + 0.018329381942749023, + 0.023358935490250587, + 0.006830048281699419, + 0.007775655947625637, + 0.0027725098188966513, + 0.011928251944482327, + 0.0108830900862813, + 0.010398570448160172, + 0.010400230064988136, + 0.009020650759339333, + 0.008648727089166641, + 0.009284330531954765, + 0.0046763173304498196, + 0.012573879212141037, + 0.012456662952899933, + 0.012958349660038948, + 0.010445320047438145, + 0.009896794334053993, + 0.006060617510229349, + 0.007063120137900114, + 0.003119488013908267, + 0.009531049989163876, + 0.011643788777291775, + 0.009153285063803196, + 0.011941660195589066, + 0.011119382455945015, + 0.004550695884972811, + 0.005657255183905363, + 0.004698129370808601, + 0.008474142290651798, + 0.009096356108784676, + 0.010139208287000656, + 0.011260001920163631, + 0.010053829289972782, + 0.006412160117179155, + 0.007280911318957806, + 0.00322563573718071, + 0.00811245758086443, + 0.0083094397559762, + 0.011584336869418621, + 0.011235080659389496, + 0.011049550026655197, + 0.004546890035271645, + 0.004344089422374964, + 0.002914699027314782, + 0.006567584350705147, + 0.0069672707468271255, + 0.0069481078535318375, + 0.008309290744364262, + 0.009698760695755482, + 0.010185806080698967, + 0.010894598439335823, + 0.007249504793435335, + 0.013685917481780052, + 0.016626251861453056, + 0.01580764167010784, + 0.017576823011040688, + 0.01376386545598507, + 0.0034808164928108454, + 0.004202219191938639, + 0.002433707704767585, + 0.003932207357138395, + 0.006482183467596769, + 0.006218153517693281, + 0.005827479064464569, + 0.005321478471159935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01606672629714012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, draws the viewer's gaze deep into the depths of a dimly lit, ancient inn. At the center of the frame, a lone swordsman sits at a greasy, square wooden table. He wears a dark blue, coarse cloth gown, slightly frayed at the collar and cuffs. His long, jet-black hair is casually tied back with a simple cloth sash, a few strands falling across his sharply defined features. A flickering oil lamp on the table provides the only light source. The warm, dim glow casts deep shadows on his face, highlighting his tightly pursed lips and eagle-like gaze. He leans forward slightly, his gaze fixed intently on the heavy wooden door not far away, as if awaiting the arrival of a fate. His right hand, with distinct joints and veins visible, grips the hilt of an ancient longsword at his waist with a gesture imbued with immense power. The hilt is made of dark wood and wrapped in aged leather. The main colors of the entire scene are dark brown and dark red, illuminated by that jumping orange light. The air is filled with a silent, focused and tense atmosphere that is about to explode, like the last calm before the storm.", + "seed": 6486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_2093_seed_6486_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010286700818687678, + "vlm_logprob": -0.010286700818687678, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.010286700818687678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010831138119101524 + ], + "opposite_logprob_tokens": [ + -4.635831356048584 + ], + "logprob_reward": -0.010831138119101524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010831138119101524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23885786533355713, + 0.2498956173658371, + 0.07862599194049835, + 0.06151162460446358, + 0.13615384697914124, + 0.02672712132334709, + 0.10132065415382385, + 0.03503935784101486, + 0.02253791317343712, + 0.049330055713653564 + ], + "credit_map": [ + 0.01821974478662014, + 0.07060567289590836, + 0.029792020097374916, + 0.014853347092866898, + 0.01579003781080246, + 0.01760399527847767, + 0.013311774469912052, + 0.058681245893239975, + 0.01762579195201397, + 0.05380774289369583, + 0.04657934233546257, + 0.016575779765844345, + 0.01595224067568779, + 0.028885100036859512, + 0.015585213899612427, + 0.0548844076693058, + 0.007596525363624096, + 0.01575210876762867, + 0.011061803437769413, + 0.004532442428171635, + 0.005427168216556311, + 0.01178957149386406, + 0.005049214698374271, + 0.01741715334355831, + 0.007765894755721092, + 0.011404394172132015, + 0.012162035331130028, + 0.0062563116662204266, + 0.0055315750651061535, + 0.005359823815524578, + 0.005182602442800999, + 0.007848984561860561, + 0.017232326790690422, + 0.022217050194740295, + 0.03486800193786621, + 0.016348373144865036, + 0.011550486087799072, + 0.011721237562596798, + 0.009547810070216656, + 0.012668547220528126, + 0.0022765432950109243, + 0.0037290260661393404, + 0.005645418073982, + 0.002793320221826434, + 0.0037054922431707382, + 0.0021505276672542095, + 0.00246005249209702, + 0.003966740798205137, + 0.011962282471358776, + 0.01840224303305149, + 0.019656799733638763, + 0.010013346560299397, + 0.00933044496923685, + 0.010414881631731987, + 0.007991389371454716, + 0.013549260795116425, + 0.0021837642416357994, + 0.009610380046069622, + 0.003767049638554454, + 0.0018303715623915195, + 0.003259577788412571, + 0.0035369666293263435, + 0.0018396793166175485, + 0.009011562913656235, + 0.0017712286207824945, + 0.005320855416357517, + 0.0017907946603372693, + 0.0012887610355392098, + 0.001842956175096333, + 0.0032957540825009346, + 0.001891172374598682, + 0.005336386617273092, + 0.004335362929850817, + 0.007436732295900583, + 0.009728476405143738, + 0.0044570607133209705, + 0.004150558263063431, + 0.009431850165128708, + 0.004105004016309977, + 0.0056850118562579155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.010831138119101524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260435398668051 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260435398668051, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260435398668051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28446558117866516, + 0.15900470316410065, + 0.07606964558362961, + 0.06772442162036896, + 0.08029443025588989, + 0.040302593261003494, + 0.1257820874452591, + 0.053012289106845856, + 0.04698030650615692, + 0.06636388599872589 + ], + "credit_map": [ + 0.024578601121902466, + 0.04639347642660141, + 0.06147913262248039, + 0.02568155899643898, + 0.038113564252853394, + 0.030475251376628876, + 0.02957959473133087, + 0.028164416551589966, + 0.01488538458943367, + 0.027123335748910904, + 0.02495759166777134, + 0.01314289029687643, + 0.01999344676733017, + 0.012827984988689423, + 0.025132881477475166, + 0.020941194146871567, + 0.007272658403962851, + 0.011851423420011997, + 0.009177369065582752, + 0.005863864440470934, + 0.010407859459519386, + 0.008007929660379887, + 0.015038230456411839, + 0.008450309745967388, + 0.007630455773323774, + 0.010052687488496304, + 0.014667829498648643, + 0.006375712808221579, + 0.006207025609910488, + 0.00798115786164999, + 0.006100586615502834, + 0.008708964101970196, + 0.008460580371320248, + 0.012102645821869373, + 0.017813967540860176, + 0.008394964039325714, + 0.006861933972686529, + 0.006896842271089554, + 0.008601133711636066, + 0.01116235926747322, + 0.0038482029922306538, + 0.005823073908686638, + 0.00820948276668787, + 0.004341383930295706, + 0.0057265423238277435, + 0.004626151639968157, + 0.004065964836627245, + 0.0036617883015424013, + 0.013068490661680698, + 0.01909627951681614, + 0.030206330120563507, + 0.014601179398596287, + 0.012691937386989594, + 0.013412924483418465, + 0.009475639089941978, + 0.013229303061962128, + 0.004645944107323885, + 0.011119800619781017, + 0.012455263175070286, + 0.005166947375983, + 0.004017995670437813, + 0.003909291699528694, + 0.004544086288660765, + 0.007152958773076534, + 0.0033302814699709415, + 0.009730949066579342, + 0.0070326621644198895, + 0.0038664836902171373, + 0.004234344698488712, + 0.0054010748863220215, + 0.0038616093806922436, + 0.009522897191345692, + 0.008133073337376118, + 0.008982040919363499, + 0.01720922626554966, + 0.0071924296207726, + 0.005718310363590717, + 0.005571712739765644, + 0.006019236519932747, + 0.007537859492003918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.005260435398668051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33275431394577026, + 0.24091525375843048, + 0.062256261706352234, + 0.04146483540534973, + 0.03549540042877197, + 0.058729082345962524, + 0.03182603046298027, + 0.048273973166942596, + 0.06568451225757599, + 0.08260037750005722 + ], + "credit_map": [ + 0.016890861093997955, + 0.09515894204378128, + 0.040644094347953796, + 0.02095893584191799, + 0.02348042093217373, + 0.020199744030833244, + 0.022409016266465187, + 0.09301229566335678, + 0.018493955954909325, + 0.04708738997578621, + 0.05501861125230789, + 0.023040378466248512, + 0.01556945126503706, + 0.01990140974521637, + 0.020400814712047577, + 0.041403237730264664, + 0.005819728597998619, + 0.015090543776750565, + 0.003932015039026737, + 0.003162035485729575, + 0.004628090187907219, + 0.006409614812582731, + 0.006646259687840939, + 0.01656796969473362, + 0.004350717645138502, + 0.006636911537498236, + 0.00804685428738594, + 0.004596049431711435, + 0.00466322572901845, + 0.0034695323556661606, + 0.00418013334274292, + 0.0055214050225913525, + 0.005612989887595177, + 0.0058340830728411674, + 0.0038598317187279463, + 0.0022116401232779026, + 0.003468085313215852, + 0.003828762099146843, + 0.004070765804499388, + 0.006609241012483835, + 0.0053029474802315235, + 0.01137568335980177, + 0.011874441057443619, + 0.0049976506270468235, + 0.008420282043516636, + 0.004202638752758503, + 0.00441282382234931, + 0.008142612874507904, + 0.0032374400179833174, + 0.006344974506646395, + 0.0043790582567453384, + 0.002389146713539958, + 0.003357725916430354, + 0.004416193347424269, + 0.0035399417392909527, + 0.004161548335105181, + 0.004551648627966642, + 0.006749233230948448, + 0.007558652199804783, + 0.0047782715409994125, + 0.005595271475613117, + 0.005523496773093939, + 0.005637786816805601, + 0.007879609242081642, + 0.0059250337071716785, + 0.010238084942102432, + 0.01561685185879469, + 0.006714289542287588, + 0.011921782046556473, + 0.004861067980527878, + 0.005424382630735636, + 0.004983014892786741, + 0.009900853037834167, + 0.012024540454149246, + 0.023361017927527428, + 0.009639907628297806, + 0.006230923347175121, + 0.009028011001646519, + 0.006276017986238003, + 0.006139107048511505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736899726092815 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736899726092815, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736899726092815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4408518970012665, + 0.09804371744394302, + 0.047463007271289825, + 0.1163879781961441, + 0.06589772552251816, + 0.025035282596945763, + 0.026728056371212006, + 0.03646870329976082, + 0.04451094567775726, + 0.09861263632774353 + ], + "credit_map": [ + 0.02559145912528038, + 0.16880665719509125, + 0.022166980430483818, + 0.014085634611546993, + 0.019904380664229393, + 0.028400832787156105, + 0.014291972853243351, + 0.14760403335094452, + 0.007363229990005493, + 0.026137027889490128, + 0.015749618411064148, + 0.007381842937320471, + 0.008651334792375565, + 0.006386617664247751, + 0.005420990753918886, + 0.020953072234988213, + 0.0042866310104727745, + 0.013551310636103153, + 0.003972148057073355, + 0.0026619250420480967, + 0.004944993648678064, + 0.00439305417239666, + 0.0038877264596521854, + 0.009765222668647766, + 0.013266941532492638, + 0.019078344106674194, + 0.020392166450619698, + 0.010823146440088749, + 0.01386688556522131, + 0.013888971880078316, + 0.010013560764491558, + 0.015057981945574284, + 0.007458210922777653, + 0.01106487400829792, + 0.01930682174861431, + 0.009107772260904312, + 0.004578526131808758, + 0.004896581172943115, + 0.004184238146990538, + 0.0053007095120847225, + 0.002037214348092675, + 0.005477319937199354, + 0.0037150071002542973, + 0.0019396855495870113, + 0.003584432415664196, + 0.0019276810344308615, + 0.0025776412803679705, + 0.003776303958147764, + 0.003522369544953108, + 0.0037329215556383133, + 0.00333942798897624, + 0.0018755493219941854, + 0.0023012622259557247, + 0.00656821159645915, + 0.0021153108682483435, + 0.0032730079255998135, + 0.0025607512798160315, + 0.009372210130095482, + 0.004755775909870863, + 0.0018762028776109219, + 0.003027422120794654, + 0.003882540389895439, + 0.002140088239684701, + 0.00885371956974268, + 0.003472772426903248, + 0.005600261501967907, + 0.010326748713850975, + 0.004097058437764645, + 0.008625694550573826, + 0.004140100907534361, + 0.004031956661492586, + 0.004216354805976152, + 0.01266148965805769, + 0.01407740917056799, + 0.028463827446103096, + 0.011559493839740753, + 0.00707524036988616, + 0.010396663099527359, + 0.0071128117851912975, + 0.00726571399718523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006736899726092815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064144670963287 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064144670963287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064144670963287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3849288821220398, + 0.15469129383563995, + 0.07475609332323074, + 0.06170399859547615, + 0.044339340180158615, + 0.04880325123667717, + 0.040638506412506104, + 0.09848642349243164, + 0.059363316744565964, + 0.03228890895843506 + ], + "credit_map": [ + 0.018087755888700485, + 0.12647387385368347, + 0.024457674473524094, + 0.014477232471108437, + 0.021773576736450195, + 0.03535817191004753, + 0.024961726740002632, + 0.11933886259794235, + 0.009653652086853981, + 0.035169873386621475, + 0.028285464271903038, + 0.0107843903824687, + 0.010753504000604153, + 0.013180690817534924, + 0.012744721956551075, + 0.03411899879574776, + 0.005302121862769127, + 0.02117929607629776, + 0.006080506835132837, + 0.003045168472453952, + 0.0051536657847464085, + 0.006622564513236284, + 0.006542936898767948, + 0.020829828456044197, + 0.007849978283047676, + 0.010269233956933022, + 0.013234782963991165, + 0.006803366355597973, + 0.005611832719296217, + 0.0056929620914161205, + 0.004819940775632858, + 0.007421898189932108, + 0.004947017878293991, + 0.0065659768879413605, + 0.011036391369998455, + 0.005703387316316366, + 0.0044952635653316975, + 0.0028012790717184544, + 0.004701866302639246, + 0.004088160581886768, + 0.005058455280959606, + 0.010313643142580986, + 0.008769538253545761, + 0.004105195868760347, + 0.006144048646092415, + 0.0047759427689015865, + 0.003505443222820759, + 0.006130984518676996, + 0.0037355467211455107, + 0.005944773089140654, + 0.011195115745067596, + 0.0045496742241084576, + 0.003596047405153513, + 0.004150516353547573, + 0.0034460844472050667, + 0.004020750056952238, + 0.007358988281339407, + 0.024204889312386513, + 0.014492608606815338, + 0.0065224310383200645, + 0.009401281364262104, + 0.007945412769913673, + 0.0055384389124810696, + 0.023022377863526344, + 0.005691461265087128, + 0.008722076192498207, + 0.014865719713270664, + 0.0064931828528642654, + 0.009964263066649437, + 0.004927559290081263, + 0.00438980758190155, + 0.004309248179197311, + 0.003851366462185979, + 0.005484253633767366, + 0.004465410020202398, + 0.002558476757258177, + 0.002856561215594411, + 0.005432795267552137, + 0.0029048046562820673, + 0.004735241178423166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016064144670963287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, draws the viewer's gaze deep into the depths of a dimly lit, ancient inn. At the center of the frame, a lone swordsman sits at a greasy, square wooden table. He wears a dark blue, coarse cloth gown, slightly frayed at the collar and cuffs. His long, jet-black hair is casually tied back with a simple cloth sash, a few strands falling across his sharply defined features. A flickering oil lamp on the table provides the only light source. The warm, dim glow casts deep shadows on his face, highlighting his tightly pursed lips and eagle-like gaze. He leans forward slightly, his gaze fixed intently on the heavy wooden door not far away, as if awaiting the arrival of a fate. His right hand, with distinct joints and veins visible, grips the hilt of an ancient longsword at his waist with a gesture imbued with immense power. The hilt is made of dark wood and wrapped in aged leather. The main colors of the entire scene are dark brown and dark red, illuminated by that jumping orange light. The air is filled with a silent, focused and tense atmosphere that is about to explode, like the last calm before the storm.", + "seed": 6487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_2093_seed_6487_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.008346628537401557, + "vlm_logprob": -0.008346628537401557, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.008346628537401557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008557554334402084 + ], + "opposite_logprob_tokens": [ + -4.883557319641113 + ], + "logprob_reward": -0.008557554334402084, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008557554334402084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11278266459703445, + 0.05838312208652496, + 0.06030019000172615, + 0.08475631475448608, + 0.1389746069908142, + 0.05102211609482765, + 0.025505810976028442, + 0.3549402356147766, + 0.06170710548758507, + 0.051627837121486664 + ], + "credit_map": [ + 0.012268402613699436, + 0.02883809432387352, + 0.017094818875193596, + 0.011707236990332603, + 0.012200860306620598, + 0.010105292312800884, + 0.008984715677797794, + 0.011583248153328896, + 0.004989818204194307, + 0.011252407915890217, + 0.009377194568514824, + 0.008754095993936062, + 0.006752707064151764, + 0.005781242623925209, + 0.004951669834554195, + 0.00652398681268096, + 0.005788142327219248, + 0.01650129072368145, + 0.008758626878261566, + 0.0052641103975474834, + 0.009125498123466969, + 0.005544104613363743, + 0.0037557005416601896, + 0.005562713835388422, + 0.009176281280815601, + 0.014534755609929562, + 0.012343686074018478, + 0.006493725348263979, + 0.008539163507521152, + 0.018371324986219406, + 0.006144022103399038, + 0.00915335863828659, + 0.028175517916679382, + 0.05532146617770195, + 0.007967260666191578, + 0.005899825599044561, + 0.008460051380097866, + 0.017926836386322975, + 0.00457354448735714, + 0.010650111362338066, + 0.005119686014950275, + 0.01134402770549059, + 0.004105766769498587, + 0.004700851626694202, + 0.005197453312575817, + 0.008906418457627296, + 0.003399420762434602, + 0.008248490281403065, + 0.002875245176255703, + 0.005036432761698961, + 0.0033399646636098623, + 0.0036369608715176582, + 0.003115083323791623, + 0.00263634929433465, + 0.002034983830526471, + 0.0028307924512773752, + 0.06889040023088455, + 0.14519204199314117, + 0.022986583411693573, + 0.01758497953414917, + 0.016373392194509506, + 0.042821697890758514, + 0.011255352757871151, + 0.029835805296897888, + 0.010111121460795403, + 0.018878871574997902, + 0.008051115088164806, + 0.004278472159057856, + 0.004465638659894466, + 0.005022105760872364, + 0.0032519951928406954, + 0.007647787686437368, + 0.007791626732796431, + 0.012787523679435253, + 0.00966356135904789, + 0.003126489231362939, + 0.004218735732138157, + 0.00587994372472167, + 0.002073143143206835, + 0.00608681607991457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.008557554334402084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744240876287222 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744240876287222, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744240876287222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07955817133188248, + 0.04143873229622841, + 0.0653013065457344, + 0.038003597408533096, + 0.2663355767726898, + 0.11313331127166748, + 0.13721971213817596, + 0.1326015591621399, + 0.03979332745075226, + 0.08661460876464844 + ], + "credit_map": [ + 0.008158537559211254, + 0.015721457079052925, + 0.011099511757493019, + 0.008293804712593555, + 0.011860748752951622, + 0.007599761709570885, + 0.007413740269839764, + 0.009410625323653221, + 0.0033356028143316507, + 0.010997476056218147, + 0.006036999169737101, + 0.005391920916736126, + 0.006271847523748875, + 0.00276716030202806, + 0.003017233684659004, + 0.0036204983480274677, + 0.00663549080491066, + 0.01680048555135727, + 0.0096264174208045, + 0.005131103098392487, + 0.01023999135941267, + 0.005772530101239681, + 0.0043152133002877235, + 0.0067800781689584255, + 0.003981328569352627, + 0.005291642155498266, + 0.0053467717953026295, + 0.00369348656386137, + 0.004595731385052204, + 0.0038303700275719166, + 0.0041287485510110855, + 0.0071355258114635944, + 0.05098811909556389, + 0.09840186685323715, + 0.014213901944458485, + 0.011792443692684174, + 0.017352253198623657, + 0.03619561344385147, + 0.011312582530081272, + 0.026078836992383003, + 0.011837013997137547, + 0.017469296231865883, + 0.01600031927227974, + 0.016428984701633453, + 0.013109118677675724, + 0.01999974623322487, + 0.008054083213210106, + 0.010234765708446503, + 0.01594296656548977, + 0.04088328033685684, + 0.013428145088255405, + 0.010799522511661053, + 0.011194851249456406, + 0.017307031899690628, + 0.010407215915620327, + 0.0172567218542099, + 0.024468516930937767, + 0.05339469760656357, + 0.008550150319933891, + 0.007260522339493036, + 0.006379290483891964, + 0.01579645834863186, + 0.005118283908814192, + 0.011633648537099361, + 0.006425380241125822, + 0.005553652066737413, + 0.007369357161223888, + 0.003980811219662428, + 0.0030860737897455692, + 0.00438621686771512, + 0.002766166813671589, + 0.006225676275789738, + 0.012270894832909107, + 0.011245903559029102, + 0.018791891634464264, + 0.004840381909161806, + 0.006111373193562031, + 0.009419554844498634, + 0.006500684656202793, + 0.017433932051062584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.006744240876287222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18859149515628815, + 0.11551999300718307, + 0.11127316206693649, + 0.04789352044463158, + 0.12962687015533447, + 0.11422265321016312, + 0.11313065141439438, + 0.08908430486917496, + 0.05128145217895508, + 0.0393759086728096 + ], + "credit_map": [ + 0.017665628343820572, + 0.04185754433274269, + 0.02622135728597641, + 0.0153647530823946, + 0.023507365956902504, + 0.02014790289103985, + 0.021271323785185814, + 0.022555600851774216, + 0.011190054006874561, + 0.023289676755666733, + 0.019407136365771294, + 0.01655242219567299, + 0.013767723925411701, + 0.010970430448651314, + 0.009451542980968952, + 0.01089099794626236, + 0.012268023565411568, + 0.027061136439442635, + 0.016357654705643654, + 0.013873339630663395, + 0.012527509592473507, + 0.00740654394030571, + 0.011182055808603764, + 0.01059688813984394, + 0.004595325794070959, + 0.013697127811610699, + 0.006135096773505211, + 0.0043584308587014675, + 0.005090826656669378, + 0.004984336905181408, + 0.0038240323774516582, + 0.005208337679505348, + 0.02608576789498329, + 0.03701841086149216, + 0.010657845064997673, + 0.007602351251989603, + 0.011355719529092312, + 0.015984468162059784, + 0.0080647561699152, + 0.012857550755143166, + 0.01090841181576252, + 0.02020350657403469, + 0.013646498322486877, + 0.01826181635260582, + 0.010224434547126293, + 0.021412312984466553, + 0.007777226157486439, + 0.011788438074290752, + 0.013734127394855022, + 0.03716982156038284, + 0.00940681528300047, + 0.008045553229749203, + 0.008698924444615841, + 0.01573815569281578, + 0.006459490861743689, + 0.013877750374376774, + 0.013542578555643559, + 0.022784888744354248, + 0.00535942567512393, + 0.0036496135871857405, + 0.0073653473518788815, + 0.018761619925498962, + 0.00670795189216733, + 0.010912870056927204, + 0.004924629349261522, + 0.013641647063195705, + 0.010019738227128983, + 0.0049798679538071156, + 0.004043659195303917, + 0.005335726775228977, + 0.003763608867302537, + 0.00457256892696023, + 0.003636640729382634, + 0.009316811338067055, + 0.003698382992297411, + 0.005755574908107519, + 0.004429427441209555, + 0.00482522789388895, + 0.0035663119051605463, + 0.004147530533373356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09619317203760147, + 0.05557303875684738, + 0.06730463355779648, + 0.025244304910302162, + 0.1061052531003952, + 0.10044655203819275, + 0.03636039420962334, + 0.39708301424980164, + 0.04220417141914368, + 0.07348547875881195 + ], + "credit_map": [ + 0.009352968074381351, + 0.018139410763978958, + 0.01391616091132164, + 0.01484959851950407, + 0.010220380499958992, + 0.008790273219347, + 0.0112224780023098, + 0.009701903909444809, + 0.005001131910830736, + 0.010952569544315338, + 0.008313062600791454, + 0.010018974542617798, + 0.007609572261571884, + 0.004387699533253908, + 0.004310013260692358, + 0.004980013240128756, + 0.006761114113032818, + 0.017601391300559044, + 0.010391674935817719, + 0.00800228025764227, + 0.008868838660418987, + 0.00586557574570179, + 0.004978415556252003, + 0.004835343919694424, + 0.0023492483887821436, + 0.005238090176135302, + 0.0034961942583322525, + 0.003108134726062417, + 0.0037744627334177494, + 0.0029545114375650883, + 0.0020543502178043127, + 0.0022693141363561153, + 0.01604580692946911, + 0.032995566725730896, + 0.008842073380947113, + 0.004742125049233437, + 0.007290654815733433, + 0.021577613428235054, + 0.0053805517964065075, + 0.009230862371623516, + 0.009593633934855461, + 0.0236359890550375, + 0.00673879636451602, + 0.01043212041258812, + 0.008484050631523132, + 0.025564009323716164, + 0.006006695330142975, + 0.009991258382797241, + 0.00438837893307209, + 0.010338199324905872, + 0.0028362853918224573, + 0.0026264297775924206, + 0.004671270493417978, + 0.004944479092955589, + 0.0025528771802783012, + 0.004002475645393133, + 0.06574883311986923, + 0.15397603809833527, + 0.027266725897789, + 0.01960618607699871, + 0.021894587203860283, + 0.0514826700091362, + 0.01618785597383976, + 0.04092007502913475, + 0.004188688937574625, + 0.010839548893272877, + 0.00533499289304018, + 0.004670355468988419, + 0.003748438786715269, + 0.005820505786687136, + 0.0034196299966424704, + 0.004182008560746908, + 0.011350429616868496, + 0.019794300198554993, + 0.01735859550535679, + 0.005592466797679663, + 0.004029963631182909, + 0.006761281285434961, + 0.003070008708164096, + 0.005528436973690987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.002818423556163907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253511756658554 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253511756658554, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253511756658554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15410387516021729, + 0.12079056352376938, + 0.11517814546823502, + 0.056762199848890305, + 0.1533585786819458, + 0.10166405141353607, + 0.048730913549661636, + 0.14955617487430573, + 0.05032126232981682, + 0.04953421279788017 + ], + "credit_map": [ + 0.013835611753165722, + 0.029596514999866486, + 0.021369801834225655, + 0.017687108367681503, + 0.020604871213436127, + 0.021378112956881523, + 0.013871630653738976, + 0.015760233625769615, + 0.009096757508814335, + 0.025432001799345016, + 0.018655454739928246, + 0.012015738524496555, + 0.015417334623634815, + 0.01127873919904232, + 0.01385397370904684, + 0.015040569007396698, + 0.009564647451043129, + 0.02678810991346836, + 0.016886863857507706, + 0.012670285999774933, + 0.013563648797571659, + 0.009294242598116398, + 0.016039885580539703, + 0.010370462201535702, + 0.005082717631012201, + 0.011189891025424004, + 0.007827866822481155, + 0.00838962011039257, + 0.007383702322840691, + 0.0047176433727145195, + 0.005452665500342846, + 0.00671809259802103, + 0.028571637347340584, + 0.049829114228487015, + 0.011283695697784424, + 0.012419351376593113, + 0.009142808616161346, + 0.018998106941580772, + 0.010381213389337063, + 0.012732649222016335, + 0.011683356016874313, + 0.02482188306748867, + 0.008377894759178162, + 0.010370221920311451, + 0.008027336560189724, + 0.01600961573421955, + 0.007105095777660608, + 0.015268649905920029, + 0.00433910358697176, + 0.009779536165297031, + 0.005185998976230621, + 0.004579848609864712, + 0.00567608280107379, + 0.007242345251142979, + 0.005179481115192175, + 0.006748514715582132, + 0.03226456791162491, + 0.05265330895781517, + 0.009541956707835197, + 0.0061557842418551445, + 0.007653497159481049, + 0.01908295415341854, + 0.006680937018245459, + 0.015523170121014118, + 0.004803176037967205, + 0.009403378702700138, + 0.009240933693945408, + 0.005394540727138519, + 0.004510314669460058, + 0.005886109080165625, + 0.004918078426271677, + 0.0061647286638617516, + 0.0050024972297251225, + 0.009817668236792088, + 0.007402488496154547, + 0.005122876726090908, + 0.004452449735254049, + 0.006641521584242582, + 0.004259766545146704, + 0.006834945175796747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.01253511756658554 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, draws the viewer's gaze deep into the depths of a dimly lit, ancient inn. At the center of the frame, a lone swordsman sits at a greasy, square wooden table. He wears a dark blue, coarse cloth gown, slightly frayed at the collar and cuffs. His long, jet-black hair is casually tied back with a simple cloth sash, a few strands falling across his sharply defined features. A flickering oil lamp on the table provides the only light source. The warm, dim glow casts deep shadows on his face, highlighting his tightly pursed lips and eagle-like gaze. He leans forward slightly, his gaze fixed intently on the heavy wooden door not far away, as if awaiting the arrival of a fate. His right hand, with distinct joints and veins visible, grips the hilt of an ancient longsword at his waist with a gesture imbued with immense power. The hilt is made of dark wood and wrapped in aged leather. The main colors of the entire scene are dark brown and dark red, illuminated by that jumping orange light. The air is filled with a silent, focused and tense atmosphere that is about to explode, like the last calm before the storm.", + "seed": 6488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_11/index_2093_seed_6488_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02720448635518551, + "vlm_logprob": -0.02720448635518551, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02720448635518551, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06485535949468613 + ], + "opposite_logprob_tokens": [ + -2.8148553371429443 + ], + "logprob_reward": -0.06485535949468613, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06485535949468613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17880740761756897, + 0.06297054886817932, + 0.055942848324775696, + 0.1999898999929428, + 0.05163794383406639, + 0.10367418825626373, + 0.058490946888923645, + 0.13519948720932007, + 0.05483078584074974, + 0.09845595061779022 + ], + "credit_map": [ + 0.015224461443722248, + 0.03346690535545349, + 0.02808712050318718, + 0.025405801832675934, + 0.027987748384475708, + 0.012865271419286728, + 0.02054419182240963, + 0.015225905925035477, + 0.004837817512452602, + 0.010877369903028011, + 0.009528475813567638, + 0.009153160266578197, + 0.009724725969135761, + 0.0052713737823069096, + 0.00792438443750143, + 0.005653242580592632, + 0.00623801164329052, + 0.014187321998178959, + 0.007798123639076948, + 0.007351358886808157, + 0.007141486741602421, + 0.003948933444917202, + 0.006123070605099201, + 0.00315454532392323, + 0.017339086160063744, + 0.03453860804438591, + 0.027689466252923012, + 0.018065230920910835, + 0.030911661684513092, + 0.0139535553753376, + 0.04350654408335686, + 0.01398575771600008, + 0.004776918329298496, + 0.011084564961493015, + 0.006759809795767069, + 0.0077312965877354145, + 0.0076242960058152676, + 0.0033822522964328527, + 0.0068735284730792046, + 0.0034052806440740824, + 0.010051075369119644, + 0.016098905354738235, + 0.01563183031976223, + 0.01085644494742155, + 0.015444155782461166, + 0.008822502568364143, + 0.01681041717529297, + 0.009958863258361816, + 0.005048672668635845, + 0.009192194789648056, + 0.00888411607593298, + 0.008131241425871849, + 0.00737953232601285, + 0.00368665368296206, + 0.012399783357977867, + 0.0037687579169869423, + 0.012745391577482224, + 0.023980487138032913, + 0.0198329146951437, + 0.01260420586913824, + 0.022775445133447647, + 0.009639039635658264, + 0.023715713992714882, + 0.009906296618282795, + 0.004505644552409649, + 0.008994380943477154, + 0.007922970689833164, + 0.005470141302794218, + 0.010184718295931816, + 0.0035172065254300833, + 0.010883416049182415, + 0.003352307714521885, + 0.0076874298974871635, + 0.014422725886106491, + 0.013497239910066128, + 0.010374314151704311, + 0.012539053335785866, + 0.007399401627480984, + 0.026239996775984764, + 0.006295786704868078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.06485535949468613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196552991867065 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196552991867065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196552991867065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08097681403160095, + 0.05695723742246628, + 0.048727408051490784, + 0.23739761114120483, + 0.14479580521583557, + 0.2906745374202728, + 0.023050200194120407, + 0.09052599966526031, + 0.01935330405831337, + 0.007541033439338207 + ], + "credit_map": [ + 0.008910889737308025, + 0.014212357811629772, + 0.009945575147867203, + 0.0113622285425663, + 0.009893170557916164, + 0.00611528754234314, + 0.014197192154824734, + 0.006340113468468189, + 0.00500286091119051, + 0.0108273820951581, + 0.007030382752418518, + 0.008161237463355064, + 0.007969387806952, + 0.004457929637283087, + 0.00815465860068798, + 0.00535339443013072, + 0.004326147027313709, + 0.006993776187300682, + 0.0060834139585494995, + 0.00540304696187377, + 0.008812326937913895, + 0.0033648007083684206, + 0.009877871721982956, + 0.0038660254795104265, + 0.02138133905827999, + 0.04379137605428696, + 0.03252695873379707, + 0.022662518545985222, + 0.03541488200426102, + 0.01642671227455139, + 0.04795825853943825, + 0.017235592007637024, + 0.015964431688189507, + 0.026999492198228836, + 0.022664953023195267, + 0.016615096479654312, + 0.024106310680508614, + 0.009815740399062634, + 0.01795758306980133, + 0.010672206059098244, + 0.028886321932077408, + 0.049988627433776855, + 0.04062280058860779, + 0.030909061431884766, + 0.04553181678056717, + 0.02341873198747635, + 0.045430198311805725, + 0.025886960327625275, + 0.0023631062358617783, + 0.0038420693017542362, + 0.0038020245265215635, + 0.0029196790419518948, + 0.00297284172847867, + 0.002009493764489889, + 0.003787307534366846, + 0.0013536786427721381, + 0.008688688278198242, + 0.016320886090397835, + 0.013303509913384914, + 0.00877122301608324, + 0.014701583422720432, + 0.006419662851840258, + 0.015622004866600037, + 0.006698437500745058, + 0.001992369769141078, + 0.004406193736940622, + 0.002524183364585042, + 0.002611567499116063, + 0.002216522814705968, + 0.001327531412243843, + 0.0025595948100090027, + 0.0017153386725112796, + 0.0006213049637153745, + 0.001303497003391385, + 0.0010022107744589448, + 0.0009550473187118769, + 0.0009906577179208398, + 0.0006390936905518174, + 0.0015206995885819197, + 0.0005085223820060492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.018196552991867065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2796231210231781, + 0.09799493849277496, + 0.07773864269256592, + 0.12260061502456665, + 0.04437093436717987, + 0.128490149974823, + 0.048536282032728195, + 0.10849431902170181, + 0.03489235043525696, + 0.05725857615470886 + ], + "credit_map": [ + 0.018883181735873222, + 0.0508170984685421, + 0.04223238304257393, + 0.04710845649242401, + 0.04580138251185417, + 0.018658999353647232, + 0.038881633430719376, + 0.017239991575479507, + 0.008604200556874275, + 0.02036282606422901, + 0.01297240424901247, + 0.013997459784150124, + 0.015975654125213623, + 0.006769427098333836, + 0.01290907897055149, + 0.006403883453458548, + 0.007785410154610872, + 0.01954362727701664, + 0.011030853725969791, + 0.011261213570833206, + 0.010498499497771263, + 0.00476826960220933, + 0.008829825557768345, + 0.004020946566015482, + 0.012031040154397488, + 0.02391945943236351, + 0.017144063487648964, + 0.01292426511645317, + 0.016564838588237762, + 0.008566563948988914, + 0.022867023944854736, + 0.008583365008234978, + 0.004406333435326815, + 0.00832432135939598, + 0.007818852551281452, + 0.006831982638686895, + 0.005999542772769928, + 0.0029259987641125917, + 0.005305073224008083, + 0.002758827991783619, + 0.012263601645827293, + 0.02116406336426735, + 0.017602235078811646, + 0.012282962910830975, + 0.02143927849829197, + 0.010173320770263672, + 0.021945154294371605, + 0.01161953341215849, + 0.004788942169398069, + 0.008045065216720104, + 0.008204764686524868, + 0.006206362042576075, + 0.005738934967666864, + 0.0034395225811749697, + 0.008995359763503075, + 0.003117334097623825, + 0.008941327221691608, + 0.01629120297729969, + 0.015551755204796791, + 0.010148659348487854, + 0.019372818991541862, + 0.00827192235738039, + 0.02197926677763462, + 0.007937364280223846, + 0.0034208991564810276, + 0.006892295088618994, + 0.004243692848831415, + 0.003917861729860306, + 0.00505549693480134, + 0.0025693601928651333, + 0.005576401017606258, + 0.003216343931853771, + 0.0040716431103646755, + 0.009917055256664753, + 0.008041736669838428, + 0.005323261953890324, + 0.01024768315255642, + 0.00415848009288311, + 0.012018189765512943, + 0.003480526152998209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736070849001408 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736070849001408, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736070849001408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20399360358715057, + 0.12556929886341095, + 0.07388446480035782, + 0.06977375596761703, + 0.08411819487810135, + 0.09877070784568787, + 0.11407992988824844, + 0.07712551206350327, + 0.10909851640462875, + 0.043586064130067825 + ], + "credit_map": [ + 0.01713172160089016, + 0.034702643752098083, + 0.027578772976994514, + 0.025087907910346985, + 0.03062036819756031, + 0.017536094412207603, + 0.030014825984835625, + 0.0213212501257658, + 0.011532978154718876, + 0.02346217818558216, + 0.015965113416314125, + 0.016514692455530167, + 0.016575928777456284, + 0.009965873323380947, + 0.017845531925559044, + 0.013706989586353302, + 0.008034982718527317, + 0.012294444255530834, + 0.01014759112149477, + 0.01028450857847929, + 0.010525709949433804, + 0.006706187501549721, + 0.009502696804702282, + 0.006388337351381779, + 0.0075659672729671, + 0.015030858106911182, + 0.009659056551754475, + 0.008073351345956326, + 0.009711941704154015, + 0.006633278448134661, + 0.008116107434034348, + 0.004983190912753344, + 0.006058620288968086, + 0.016043715178966522, + 0.011556646786630154, + 0.009455895982682705, + 0.010500412434339523, + 0.006807703524827957, + 0.015957532450556755, + 0.007737660314887762, + 0.010438460856676102, + 0.020772868767380714, + 0.013078806921839714, + 0.013517041690647602, + 0.018117843195796013, + 0.0049804095178842545, + 0.011689416132867336, + 0.006175846792757511, + 0.012261827476322651, + 0.025301240384578705, + 0.016104286536574364, + 0.014340751804411411, + 0.014079099521040916, + 0.007315215189009905, + 0.014475592412054539, + 0.010201903991401196, + 0.007407248020172119, + 0.013954119756817818, + 0.01070769876241684, + 0.008704136125743389, + 0.013076234608888626, + 0.0055503020994365215, + 0.011298197321593761, + 0.0064275613985955715, + 0.0101550268009305, + 0.02134561538696289, + 0.013714253902435303, + 0.011163091287016869, + 0.015524100512266159, + 0.008874129503965378, + 0.01742166467010975, + 0.010900615714490414, + 0.0038342932239174843, + 0.004987697117030621, + 0.0052254158072173595, + 0.006491636857390404, + 0.00631176121532917, + 0.003904603188857436, + 0.009614606387913227, + 0.0032160466071218252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.006736070849001408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, ancient inn with a heavy wooden door visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone swordsman sitting at a greasy, square wooden table in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swordsman wear a dark blue, coarse cloth gown that is slightly frayed at the collar and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by a flickering oil lamp on the table, casting deep shadows and showing dark brown and dark red main colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, taken from a level angle, and does the swordsman grip an ancient longsword hilt made of dark wood wrapped in aged leather?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2741715610027313, + 0.11524598300457001, + 0.06706774234771729, + 0.06343094259500504, + 0.06233831122517586, + 0.057781852781772614, + 0.03419165313243866, + 0.12035121023654938, + 0.0893465206027031, + 0.11607427150011063 + ], + "credit_map": [ + 0.026204586029052734, + 0.05041082948446274, + 0.03140438720583916, + 0.0348653718829155, + 0.03915153816342354, + 0.022360771894454956, + 0.04995756223797798, + 0.019816488027572632, + 0.009676705114543438, + 0.012927099131047726, + 0.016083138063549995, + 0.013252208940684795, + 0.015155285596847534, + 0.008602705784142017, + 0.03228600323200226, + 0.007262820843607187, + 0.005757717415690422, + 0.008446980267763138, + 0.008408094756305218, + 0.010347051545977592, + 0.009037927724421024, + 0.004658517427742481, + 0.016435910016298294, + 0.003975535277277231, + 0.004049171693623066, + 0.007034613750874996, + 0.009153619408607483, + 0.010726101696491241, + 0.007993683218955994, + 0.0034232367761433125, + 0.016290754079818726, + 0.004759758710861206, + 0.005668327212333679, + 0.009817429818212986, + 0.008664757944643497, + 0.010057972744107246, + 0.007235589437186718, + 0.004240743350237608, + 0.012557988055050373, + 0.0040954952128231525, + 0.00477147102355957, + 0.008899680338799953, + 0.007889159955084324, + 0.00676826061680913, + 0.01041932962834835, + 0.0042227027006447315, + 0.009641737677156925, + 0.005169508513063192, + 0.003352154977619648, + 0.0049922107718884945, + 0.004890070762485266, + 0.004257132299244404, + 0.004172167275100946, + 0.0031800665892660618, + 0.007060573436319828, + 0.0022872742265462875, + 0.01024652924388647, + 0.02058149315416813, + 0.01586846262216568, + 0.010583375580608845, + 0.022629564628005028, + 0.01006113551557064, + 0.019530193880200386, + 0.010850450955331326, + 0.008932417258620262, + 0.019367940723896027, + 0.01145852729678154, + 0.009815004654228687, + 0.012047979049384594, + 0.006616631522774696, + 0.013011067174375057, + 0.00809694267809391, + 0.008447171188890934, + 0.01640513725578785, + 0.01603269763290882, + 0.011159569956362247, + 0.017718138173222542, + 0.009321186691522598, + 0.0293519739061594, + 0.007638383191078901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.01253888476639986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_tensor([3408, 3408])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_tensor([3408, 3408])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bda4ee2d86469b9931042f0c045674d31f54b473 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_tensor([3408, 3408])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme slow-motion shot shot from the ground, the world is reimagined as a child's drawing, brimming with playful innocence. At the center of the frame lies a forest of purple-lined, green-skinned bamboo. Each bamboo pole is not a perfect cylinder, but rather a slightly skewed, crayon-like line drawn with childish strokes. The color is unevenly filled, revealing a paper-like texture. Striking vertical purple stripes are meticulously inscribed on these apple-green poles, their edges tinged with the characteristic roughness of crayon. In the almost static flow of time, the bamboo forest sways with a dreamlike, almost imperceptible movement. The bamboo leaves, simplified into dark green paper-cut shapes, tremble subtly with the swaying of the poles. Most strikingly, with each subtle bend of the poles, the purple stripes appear to slowly expand and contract, creating a sense of life, akin to \"breathing.\" The background is a flat, soft ochre earth and azure sky, dotted with a few fluffy white clouds like cotton candy. In the corner, a huge yellow sun with a smiling face casts warm and even light, injecting infinite tenderness and reverie into this quiet, innocent fairy tale scene.", + "seed": 21420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_3408_seed_21420_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012662954535335303, + "vlm_logprob": -0.012662954535335303, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.012662954535335303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010085219517350197 + ], + "opposite_logprob_tokens": [ + -4.635085105895996 + ], + "logprob_reward": -0.010085219517350197, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010085219517350197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16969670355319977, + 0.12558436393737793, + 0.09028193354606628, + 0.037640661001205444, + 0.05108160153031349, + 0.3139811158180237, + 0.08876395970582962, + 0.0341353714466095, + 0.04500927776098251, + 0.04382500797510147 + ], + "credit_map": [ + 0.024490641430020332, + 0.010262965224683285, + 0.029773790389299393, + 0.04076049476861954, + 0.006119291763752699, + 0.009088905528187752, + 0.03903144225478172, + 0.010169179178774357, + 0.01858590357005596, + 0.006109945010393858, + 0.04410605877637863, + 0.018355289474129677, + 0.0037945087533444166, + 0.0049989959225058556, + 0.023165306076407433, + 0.0064683519303798676, + 0.01235908456146717, + 0.00548657076433301, + 0.011805579997599125, + 0.019432369619607925, + 0.004924459382891655, + 0.007793902885168791, + 0.019085567444562912, + 0.009394404478371143, + 0.00524044269695878, + 0.0032554655335843563, + 0.0037212157621979713, + 0.007879480719566345, + 0.0019197765504941344, + 0.0032088113948702812, + 0.0063273729756474495, + 0.006088096182793379, + 0.007063256110996008, + 0.004011920187622309, + 0.006957272067666054, + 0.013016357086598873, + 0.0019035900477319956, + 0.0031595355831086636, + 0.0103428540751338, + 0.004626814741641283, + 0.014295142143964767, + 0.01219182088971138, + 0.021019751206040382, + 0.07074011862277985, + 0.011256726458668709, + 0.010668979957699776, + 0.10781195759773254, + 0.06599664688110352, + 0.011189697310328484, + 0.009745999239385128, + 0.00845995731651783, + 0.01495918445289135, + 0.006874230224639177, + 0.01299433596432209, + 0.012843089178204536, + 0.011697468347847462, + 0.004416958894580603, + 0.0027712101582437754, + 0.005172980483621359, + 0.006067701615393162, + 0.002094509080052376, + 0.0035282960161566734, + 0.0065230634063482285, + 0.0035606499295681715, + 0.004837928805500269, + 0.0034689996391534805, + 0.007707028184086084, + 0.008006136864423752, + 0.00311180274002254, + 0.0031736050732433796, + 0.009742645546793938, + 0.004961130674928427, + 0.006599706131964922, + 0.004327954258769751, + 0.008058481849730015, + 0.00798099860548973, + 0.002838371554389596, + 0.002656156662851572, + 0.005861982703208923, + 0.005501359235495329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.010085219517350197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06698776036500931, + 0.027862193062901497, + 0.02464395947754383, + 0.05881660059094429, + 0.13485930860042572, + 0.5886982083320618, + 0.03683362901210785, + 0.011270752176642418, + 0.016007019206881523, + 0.034020647406578064 + ], + "credit_map": [ + 0.01308569498360157, + 0.004998404998332262, + 0.013855409808456898, + 0.013493038713932037, + 0.002830703044310212, + 0.004116475582122803, + 0.00956813246011734, + 0.005039888899773359, + 0.0036406521685421467, + 0.0024272100999951363, + 0.007543288171291351, + 0.0037438347935676575, + 0.0013755391119048, + 0.0023782129865139723, + 0.00459437258541584, + 0.0021590793039649725, + 0.00443044351413846, + 0.0019013972487300634, + 0.004047737456858158, + 0.004033590201288462, + 0.0012733712792396545, + 0.003039053175598383, + 0.0034362629521638155, + 0.0024821022525429726, + 0.007358600851148367, + 0.0040696728974580765, + 0.006493480876088142, + 0.014480305835604668, + 0.0030536821577697992, + 0.004677795805037022, + 0.011858395300805569, + 0.006824658717960119, + 0.009642267599701881, + 0.00839837547391653, + 0.007647149730473757, + 0.041088249534368515, + 0.007292310241609812, + 0.009764546528458595, + 0.039775408804416656, + 0.011250992305576801, + 0.022914590314030647, + 0.027368519455194473, + 0.03984910249710083, + 0.07057302445173264, + 0.022138621658086777, + 0.03185495361685753, + 0.19211135804653168, + 0.18188796937465668, + 0.003461851505562663, + 0.00396178662776947, + 0.005416007246822119, + 0.005988442804664373, + 0.0020376378670334816, + 0.004276666324585676, + 0.005090336315333843, + 0.006600894033908844, + 0.0012826410820707679, + 0.0010532665764912963, + 0.0021408572793006897, + 0.002524408046156168, + 0.0008609115611761808, + 0.0009482252062298357, + 0.0014175391988828778, + 0.001042902353219688, + 0.0018021221039816737, + 0.001639737281948328, + 0.0030048727057874203, + 0.0032196305692195892, + 0.0011498982785269618, + 0.0014547997852787375, + 0.002228202298283577, + 0.0015077526913955808, + 0.003265985054895282, + 0.0029189849738031626, + 0.0038620743434876204, + 0.009422850795090199, + 0.0022192583419382572, + 0.0025681997649371624, + 0.005014957394450903, + 0.004748329520225525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.026350026950240135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04181844741106033, + 0.024493983015418053, + 0.04111142456531525, + 0.033087898045778275, + 0.02985367178916931, + 0.7411447763442993, + 0.04521063342690468, + 0.016886718571186066, + 0.010704383254051208, + 0.01568804308772087 + ], + "credit_map": [ + 0.005238895770162344, + 0.0035247551277279854, + 0.008779937401413918, + 0.008536095730960369, + 0.002174678258597851, + 0.003741969121620059, + 0.006585027556866407, + 0.003237085184082389, + 0.0029531759209930897, + 0.0025809381622821093, + 0.003878592746332288, + 0.004093446768820286, + 0.0014168835477903485, + 0.0022221426479518414, + 0.004652916919440031, + 0.0026958866510540247, + 0.00669010728597641, + 0.002315209712833166, + 0.006152379792183638, + 0.010312514379620552, + 0.001937077147886157, + 0.002109023043885827, + 0.009340576827526093, + 0.002254536608234048, + 0.003138533793389797, + 0.0023911905009299517, + 0.003111287485808134, + 0.008813566528260708, + 0.0017978077521547675, + 0.0021503139287233353, + 0.007937295362353325, + 0.0037479018792510033, + 0.003180217696353793, + 0.004208951722830534, + 0.003247939981520176, + 0.0044121043756604195, + 0.0025275524239987135, + 0.003974020481109619, + 0.00427114125341177, + 0.004031742457300425, + 0.030282383784651756, + 0.027423474937677383, + 0.04897760599851608, + 0.1510968655347824, + 0.026463570073246956, + 0.023615581914782524, + 0.2652008831501007, + 0.16808435320854187, + 0.006991350091993809, + 0.004657216835767031, + 0.005319440737366676, + 0.009316377341747284, + 0.0025612805038690567, + 0.005448407959192991, + 0.00568865891546011, + 0.0052279019728302956, + 0.002636242425069213, + 0.001186156878247857, + 0.0030286130495369434, + 0.004023893270641565, + 0.0010843720519915223, + 0.0013645011931657791, + 0.0024172402918338776, + 0.0011456990614533424, + 0.0014773127622902393, + 0.0008076075464487076, + 0.001977388048544526, + 0.00246230885386467, + 0.000637311371974647, + 0.0008817524649202824, + 0.0015859456034377217, + 0.0008747570100240409, + 0.0015083322068676353, + 0.0009468663483858109, + 0.002260052366182208, + 0.0031682627741247416, + 0.0009495852864347398, + 0.0014570672065019608, + 0.002815959742292762, + 0.0025819172151386738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.006732400041073561 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088377580046654 + ], + "opposite_logprob_tokens": [ + -5.504088401794434 + ], + "logprob_reward": -0.004088377580046654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088377580046654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16262753307819366, + 0.053689513355493546, + 0.18878065049648285, + 0.037505991756916046, + 0.02634299360215664, + 0.14307081699371338, + 0.14330779016017914, + 0.18381860852241516, + 0.03534295782446861, + 0.025513168424367905 + ], + "credit_map": [ + 0.030206162482500076, + 0.012205662205815315, + 0.04413183778524399, + 0.02818469889461994, + 0.0065461122430861, + 0.009167906828224659, + 0.022315634414553642, + 0.009869509376585484, + 0.00893914233893156, + 0.0030857070814818144, + 0.02012673392891884, + 0.007837978191673756, + 0.0018416354432702065, + 0.0024997820146381855, + 0.005671819671988487, + 0.0036867130547761917, + 0.03316652774810791, + 0.01234845444560051, + 0.030247826129198074, + 0.039547987282276154, + 0.00828887801617384, + 0.01045338623225689, + 0.04005758836865425, + 0.014670003205537796, + 0.003055369947105646, + 0.0025556995533406734, + 0.002844309899955988, + 0.012055660597980022, + 0.0016599235823377967, + 0.0018579225288704038, + 0.010526587255299091, + 0.0029505190905183554, + 0.003318804781883955, + 0.0028376327827572823, + 0.0039507076144218445, + 0.005045559722930193, + 0.0016256419476121664, + 0.0028928047977387905, + 0.0037203507963567972, + 0.002951489994302392, + 0.01302011776715517, + 0.010216682218015194, + 0.01466416660696268, + 0.048580195754766464, + 0.004502383526414633, + 0.009668698534369469, + 0.016895649954676628, + 0.025522928684949875, + 0.018666615709662437, + 0.01322938036173582, + 0.014467660337686539, + 0.02490108087658882, + 0.009661934338510036, + 0.019560031592845917, + 0.02135419100522995, + 0.021466897800564766, + 0.018819961696863174, + 0.012363752350211143, + 0.030793679878115654, + 0.048806123435497284, + 0.010100034065544605, + 0.011994834989309311, + 0.034998077899217606, + 0.015942160040140152, + 0.003839240875095129, + 0.002711510518565774, + 0.006611171644181013, + 0.005795573815703392, + 0.0026866223197430372, + 0.003458046820014715, + 0.006439493503421545, + 0.0038012973964214325, + 0.0031757864635437727, + 0.003116439562290907, + 0.004665610380470753, + 0.0042487154714763165, + 0.0018269213614985347, + 0.0017228961223736405, + 0.002981726545840502, + 0.003775073681026697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.004088377580046654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058748587965965 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058748587965965, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058748587965965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2656160891056061, + 0.12672927975654602, + 0.07552878558635712, + 0.03613661602139473, + 0.0310894176363945, + 0.2085559070110321, + 0.07574182748794556, + 0.08333852142095566, + 0.043297987431287766, + 0.05396561697125435 + ], + "credit_map": [ + 0.04688766971230507, + 0.017053276300430298, + 0.0876009613275528, + 0.04605108126997948, + 0.009359685704112053, + 0.013729434460401535, + 0.031109921634197235, + 0.013824052177369595, + 0.023300612345337868, + 0.0076621039770543575, + 0.031093815341591835, + 0.019844111055135727, + 0.005143134389072657, + 0.006374146323651075, + 0.026991114020347595, + 0.00632023299112916, + 0.011982248164713383, + 0.0045780823566019535, + 0.0102023771032691, + 0.015053454786539078, + 0.003632503328844905, + 0.0038556677754968405, + 0.021085696294903755, + 0.005138747859746218, + 0.0029658630955964327, + 0.002127249026671052, + 0.0027835259679704905, + 0.01158690731972456, + 0.0015364999417215586, + 0.0023292277473956347, + 0.010074627585709095, + 0.002732711611315608, + 0.0038832142017781734, + 0.0026851315051317215, + 0.00460420735180378, + 0.006887899246066809, + 0.0021357075311243534, + 0.0030060771387070417, + 0.004880510736256838, + 0.0030066680628806353, + 0.011785692535340786, + 0.008719312027096748, + 0.016333799809217453, + 0.05300050973892212, + 0.0070613292045891285, + 0.005990105681121349, + 0.0697251409292221, + 0.03593998774886131, + 0.009935955516994, + 0.007482793182134628, + 0.008966183289885521, + 0.020331602543592453, + 0.004138574004173279, + 0.006581517867743969, + 0.01145654171705246, + 0.006848648190498352, + 0.009469942189753056, + 0.007855691947042942, + 0.018931791186332703, + 0.017983514815568924, + 0.004624314606189728, + 0.004883302375674248, + 0.012479417026042938, + 0.007110531907528639, + 0.004808247089385986, + 0.0026256295386701822, + 0.010624702088534832, + 0.008860204368829727, + 0.002754389541223645, + 0.0037338868714869022, + 0.006319844163954258, + 0.0035710788797587156, + 0.005683762952685356, + 0.005045708268880844, + 0.004811524413526058, + 0.017227046191692352, + 0.003944647964090109, + 0.0048324610106647015, + 0.006956076715141535, + 0.005464379675686359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016058748587965965 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot shot from the ground, the world is reimagined as a child's drawing, brimming with playful innocence. At the center of the frame lies a forest of purple-lined, green-skinned bamboo. Each bamboo pole is not a perfect cylinder, but rather a slightly skewed, crayon-like line drawn with childish strokes. The color is unevenly filled, revealing a paper-like texture. Striking vertical purple stripes are meticulously inscribed on these apple-green poles, their edges tinged with the characteristic roughness of crayon. In the almost static flow of time, the bamboo forest sways with a dreamlike, almost imperceptible movement. The bamboo leaves, simplified into dark green paper-cut shapes, tremble subtly with the swaying of the poles. Most strikingly, with each subtle bend of the poles, the purple stripes appear to slowly expand and contract, creating a sense of life, akin to \"breathing.\" The background is a flat, soft ochre earth and azure sky, dotted with a few fluffy white clouds like cotton candy. In the corner, a huge yellow sun with a smiling face casts warm and even light, injecting infinite tenderness and reverie into this quiet, innocent fairy tale scene.", + "seed": 21421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_3408_seed_21421_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012304250476881861, + "vlm_logprob": -0.012304250476881861, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.012304250476881861, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006255806889384985 + ], + "opposite_logprob_tokens": [ + -5.131255626678467 + ], + "logprob_reward": -0.006255806889384985, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006255806889384985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5382625460624695, + 0.14106419682502747, + 0.040487051010131836, + 0.03622810170054436, + 0.021555421873927116, + 0.0324397087097168, + 0.10232704132795334, + 0.01730717159807682, + 0.04984903335571289, + 0.02047971822321415 + ], + "credit_map": [ + 0.029589569196105003, + 0.03271596506237984, + 0.04622700437903404, + 0.1956745833158493, + 0.0713929608464241, + 0.028100989758968353, + 0.05257182940840721, + 0.08198967576026917, + 0.00810105912387371, + 0.008801049552857876, + 0.01165708526968956, + 0.04790908098220825, + 0.012497778050601482, + 0.010442734695971012, + 0.01597258448600769, + 0.025682836771011353, + 0.004737904295325279, + 0.004998418502509594, + 0.0042794207111001015, + 0.0035810580011457205, + 0.00859778095036745, + 0.005906723439693451, + 0.0029277584981173277, + 0.005457988940179348, + 0.004280760418623686, + 0.005215799435973167, + 0.003939412534236908, + 0.005589330568909645, + 0.0035670807119458914, + 0.005694786086678505, + 0.004196037072688341, + 0.0037449002265930176, + 0.003341224743053317, + 0.0020533674396574497, + 0.0023417645134031773, + 0.002687381813302636, + 0.002747141756117344, + 0.002009268617257476, + 0.003328584833070636, + 0.00304669002071023, + 0.004316786769777536, + 0.0023750034160912037, + 0.008725618943572044, + 0.002278262982144952, + 0.0028156922198832035, + 0.003125555580481887, + 0.0020425263792276382, + 0.00676026800647378, + 0.007891071029007435, + 0.005869750399142504, + 0.014612599276006222, + 0.010952435433864594, + 0.01139973197132349, + 0.006900318898260593, + 0.006265741307288408, + 0.038435399532318115, + 0.002946785418316722, + 0.0023905530106276274, + 0.0020769324619323015, + 0.0011674683773890138, + 0.0018140152096748352, + 0.002854119287803769, + 0.002056616125628352, + 0.0020006836857646704, + 0.009170962497591972, + 0.0034402869641780853, + 0.014141461811959743, + 0.004220294766128063, + 0.003222979139536619, + 0.004290522541850805, + 0.0035516726784408092, + 0.007810859475284815, + 0.0037543419748544693, + 0.002443519188091159, + 0.002554133767262101, + 0.001970241079106927, + 0.0021318166982382536, + 0.0026445426046848297, + 0.002136624651029706, + 0.0028445005882531404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.006255806889384985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636047434061766 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636047434061766, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636047434061766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3940517008304596, + 0.22580328583717346, + 0.08872289210557938, + 0.08919385075569153, + 0.027706822380423546, + 0.029203983023762703, + 0.07127970457077026, + 0.029377149417996407, + 0.029389025643467903, + 0.015271523036062717 + ], + "credit_map": [ + 0.014588207937777042, + 0.018382055684924126, + 0.030316321179270744, + 0.15378475189208984, + 0.03841765224933624, + 0.018121011555194855, + 0.03469642251729965, + 0.08574530482292175, + 0.012846230529248714, + 0.014458712190389633, + 0.02199641987681389, + 0.07244265079498291, + 0.04477684572339058, + 0.017861558124423027, + 0.02780279330909252, + 0.013618078082799911, + 0.008360929787158966, + 0.006227332632988691, + 0.011781875975430012, + 0.025214795023202896, + 0.013974960893392563, + 0.006724604871124029, + 0.008692131377756596, + 0.007746261078864336, + 0.012395109981298447, + 0.00924965925514698, + 0.010230593383312225, + 0.007256379816681147, + 0.01451902650296688, + 0.014755291864275932, + 0.012676619924604893, + 0.008111173287034035, + 0.005382015369832516, + 0.002351393224671483, + 0.004432559944689274, + 0.002844171831384301, + 0.0022932637948542833, + 0.0026663271710276604, + 0.0037058042362332344, + 0.004031289368867874, + 0.004370968323200941, + 0.002446152037009597, + 0.0069009168073534966, + 0.002123106038197875, + 0.007033717352896929, + 0.003106364281848073, + 0.0017677490832284093, + 0.0014550071209669113, + 0.005042647011578083, + 0.004359008744359016, + 0.012373688630759716, + 0.002807688433676958, + 0.01082331221550703, + 0.004757365211844444, + 0.005875723902136087, + 0.025240270420908928, + 0.004100305028259754, + 0.0031196740455925465, + 0.0061127496883273125, + 0.002337230136618018, + 0.0036360095255076885, + 0.0036078253760933876, + 0.0035232314839959145, + 0.0029401243664324284, + 0.003855051938444376, + 0.0020997244864702225, + 0.004967420361936092, + 0.0035709538497030735, + 0.00169077527243644, + 0.0022065213415771723, + 0.003698361571878195, + 0.007300214841961861, + 0.0022608379367738962, + 0.001584445359185338, + 0.001961387461051345, + 0.0016433063428848982, + 0.001464860630221665, + 0.0016468114918097854, + 0.002002547960728407, + 0.0027073263190686703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.007636047434061766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036109976936131716 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036109976936131716, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036109976936131716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18492674827575684, + 0.4565203785896301, + 0.04308074712753296, + 0.043453704565763474, + 0.04194493591785431, + 0.07888340950012207, + 0.08527348935604095, + 0.021771233528852463, + 0.019228266552090645, + 0.024917103350162506 + ], + "credit_map": [ + 0.016049176454544067, + 0.014293906278908253, + 0.014168488793075085, + 0.05297943204641342, + 0.011449023149907589, + 0.014665353111922741, + 0.016409756615757942, + 0.04491160064935684, + 0.020608576014637947, + 0.028581637889146805, + 0.04253359138965607, + 0.16860634088516235, + 0.05804905667901039, + 0.0266819316893816, + 0.04458251968026161, + 0.06687674671411514, + 0.003739136504009366, + 0.00650453707203269, + 0.0033230369444936514, + 0.004339627455919981, + 0.006171379704028368, + 0.005794452503323555, + 0.006065638270229101, + 0.007142937276512384, + 0.0060348184779286385, + 0.005260927136987448, + 0.003934785258024931, + 0.005183679051697254, + 0.006378939375281334, + 0.004065306857228279, + 0.006172561086714268, + 0.00642268406227231, + 0.007178970612585545, + 0.0031014960259199142, + 0.004894052632153034, + 0.0040516178123652935, + 0.005991479381918907, + 0.0033200285397469997, + 0.00564922159537673, + 0.007758069317787886, + 0.007495713885873556, + 0.006949440110474825, + 0.017437536269426346, + 0.004333771299570799, + 0.02317376434803009, + 0.010133401490747929, + 0.004934420809149742, + 0.004425350576639175, + 0.010226850397884846, + 0.0041747018694877625, + 0.025166738778352737, + 0.005536203272640705, + 0.0031186731066554785, + 0.0058135599829256535, + 0.005628115497529507, + 0.025608645752072334, + 0.0038295420818030834, + 0.0021321445237845182, + 0.002213916042819619, + 0.002815734129399061, + 0.0020012527238577604, + 0.0024507094640284777, + 0.0021361149847507477, + 0.004191819578409195, + 0.0034995279274880886, + 0.0015054766554385424, + 0.0029750121757388115, + 0.003132777288556099, + 0.0017061647959053516, + 0.0022465793881565332, + 0.002808711025863886, + 0.001354015083052218, + 0.004907532595098019, + 0.0023468569852411747, + 0.002861684886738658, + 0.002221093513071537, + 0.003574537578970194, + 0.0033582416363060474, + 0.0015847983304411173, + 0.004062356892973185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0036109976936131716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005939689930528402 + ], + "opposite_logprob_tokens": [ + -5.130939483642578 + ], + "logprob_reward": -0.005939689930528402, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005939689930528402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6729380488395691, + 0.09970927238464355, + 0.021705837920308113, + 0.10848090052604675, + 0.005417067091912031, + 0.03270435705780983, + 0.01749463379383087, + 0.017519352957606316, + 0.018067628145217896, + 0.005962822120636702 + ], + "credit_map": [ + 0.021374965086579323, + 0.03619405999779701, + 0.0569232739508152, + 0.2564091980457306, + 0.1162932813167572, + 0.03277307376265526, + 0.06967173516750336, + 0.08329848945140839, + 0.006613555829972029, + 0.007171106990426779, + 0.010892982594668865, + 0.03027823381125927, + 0.020937826484441757, + 0.0070261782966554165, + 0.011526971124112606, + 0.005262415856122971, + 0.002541907597333193, + 0.003974129445850849, + 0.001976698637008667, + 0.004055467899888754, + 0.0027628832031041384, + 0.0028853111434727907, + 0.0019009590614587069, + 0.0016084800008684397, + 0.013043582439422607, + 0.020266883075237274, + 0.010324566625058651, + 0.013184555806219578, + 0.013791400007903576, + 0.01696145161986351, + 0.011631294153630733, + 0.0092771602794528, + 0.0009267014102078974, + 0.0004812258994206786, + 0.0008802265510894358, + 0.00045669529936276376, + 0.00054232805268839, + 0.0005497843376360834, + 0.0006294016493484378, + 0.0009507034556008875, + 0.002721911296248436, + 0.0028389417566359043, + 0.008403823710978031, + 0.0018571233376860619, + 0.009844417683780193, + 0.004511650651693344, + 0.0016066462267190218, + 0.0009198448969982564, + 0.0020273523405194283, + 0.001321157906204462, + 0.0015689108986407518, + 0.004823130555450916, + 0.0006312557379715145, + 0.0011186811607331038, + 0.0009673101012594998, + 0.0050368355587124825, + 0.003937985748052597, + 0.00179814372677356, + 0.004179148003458977, + 0.0017474545165896416, + 0.0013308824272826314, + 0.0016068757977336645, + 0.0013621581019833684, + 0.0015567051013931632, + 0.002645908622071147, + 0.0013298697303980589, + 0.0020106693264096975, + 0.0015243224333971739, + 0.0026262400206178427, + 0.0017531816847622395, + 0.0027351875323802233, + 0.0034422471653670073, + 0.0015405912417918444, + 0.0005336229805834591, + 0.0012494144029915333, + 0.0005369418067857623, + 0.0006107602966949344, + 0.0005854343180544674, + 0.0004236983950249851, + 0.0004823584749829024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005939689930528402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038078710436820984 + ], + "opposite_logprob_tokens": [ + -3.288078784942627 + ], + "logprob_reward": -0.038078710436820984, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038078710436820984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4855369031429291, + 0.059500835835933685, + 0.024701666086912155, + 0.021219896152615547, + 0.038752760738134384, + 0.08748951554298401, + 0.07690797001123428, + 0.10770794004201889, + 0.03905342146754265, + 0.05912904441356659 + ], + "credit_map": [ + 0.017409544438123703, + 0.02762182056903839, + 0.040583182126283646, + 0.18029063940048218, + 0.08521926403045654, + 0.0243549682199955, + 0.04726995900273323, + 0.06278753280639648, + 0.004366768989712, + 0.003901072544977069, + 0.0060658808797597885, + 0.01350465789437294, + 0.01319038588553667, + 0.004668287932872772, + 0.006989868823438883, + 0.0068139079958200455, + 0.003256299765780568, + 0.0028751734644174576, + 0.0030960310250520706, + 0.002795860171318054, + 0.003945495467633009, + 0.0032579819671809673, + 0.0027060939464718103, + 0.002768728183582425, + 0.0024252061266452074, + 0.0025682563427835703, + 0.0024177159648388624, + 0.0026868204586207867, + 0.001994313905015588, + 0.002993634669110179, + 0.002878595609217882, + 0.0032553519122302532, + 0.007380611728876829, + 0.004466861952096224, + 0.006030939053744078, + 0.004985104780644178, + 0.00489448057487607, + 0.003110901452600956, + 0.004677451215684414, + 0.0032064123079180717, + 0.007833986543118954, + 0.007581130601465702, + 0.021267730742692947, + 0.006237890105694532, + 0.024767696857452393, + 0.012351715937256813, + 0.004520168527960777, + 0.0029291994869709015, + 0.012211530469357967, + 0.005608873441815376, + 0.01472153328359127, + 0.007318783085793257, + 0.004735099151730537, + 0.005322131793946028, + 0.0036956914700567722, + 0.02329433150589466, + 0.02855684421956539, + 0.010774333029985428, + 0.025978917255997658, + 0.0072288597002625465, + 0.007111185230314732, + 0.009107392281293869, + 0.005345685873180628, + 0.013604721054434776, + 0.006289862561970949, + 0.007462832611054182, + 0.0035182565916329622, + 0.002695478731766343, + 0.007909612730145454, + 0.006136943586170673, + 0.0024961987510323524, + 0.002544238930568099, + 0.01630794256925583, + 0.0054852645844221115, + 0.016610877588391304, + 0.0031882040202617645, + 0.003293439745903015, + 0.0050221215933561325, + 0.0028298457618802786, + 0.006391345057636499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.038078710436820984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot shot from the ground, the world is reimagined as a child's drawing, brimming with playful innocence. At the center of the frame lies a forest of purple-lined, green-skinned bamboo. Each bamboo pole is not a perfect cylinder, but rather a slightly skewed, crayon-like line drawn with childish strokes. The color is unevenly filled, revealing a paper-like texture. Striking vertical purple stripes are meticulously inscribed on these apple-green poles, their edges tinged with the characteristic roughness of crayon. In the almost static flow of time, the bamboo forest sways with a dreamlike, almost imperceptible movement. The bamboo leaves, simplified into dark green paper-cut shapes, tremble subtly with the swaying of the poles. Most strikingly, with each subtle bend of the poles, the purple stripes appear to slowly expand and contract, creating a sense of life, akin to \"breathing.\" The background is a flat, soft ochre earth and azure sky, dotted with a few fluffy white clouds like cotton candy. In the corner, a huge yellow sun with a smiling face casts warm and even light, injecting infinite tenderness and reverie into this quiet, innocent fairy tale scene.", + "seed": 21422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_3408_seed_21422_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018703245371580125, + "vlm_logprob": -0.018703245371580125, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.018703245371580125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020980216562747955 + ], + "opposite_logprob_tokens": [ + -3.8959801197052 + ], + "logprob_reward": -0.020980216562747955, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020980216562747955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31561267375946045, + 0.07258985936641693, + 0.014047686010599136, + 0.11559804528951645, + 0.03628682345151901, + 0.03010248765349388, + 0.04066256433725357, + 0.1788375973701477, + 0.13205288350582123, + 0.06420939415693283 + ], + "credit_map": [ + 0.034040309488773346, + 0.013104815036058426, + 0.10241366177797318, + 0.025577593594789505, + 0.01692225970327854, + 0.017847757786512375, + 0.07835313677787781, + 0.02735316939651966, + 0.009587391279637814, + 0.006575893145054579, + 0.01824607141315937, + 0.00677422434091568, + 0.006581134162843227, + 0.006684455554932356, + 0.011304489336907864, + 0.006836199667304754, + 0.0018702386878430843, + 0.001550016226246953, + 0.0019477723399177194, + 0.0026245717890560627, + 0.001383068854920566, + 0.0013539487263187766, + 0.002064004773274064, + 0.0012540643801912665, + 0.015554316341876984, + 0.019747663289308548, + 0.0192584078758955, + 0.014091374352574348, + 0.011130006052553654, + 0.012914166785776615, + 0.01146209891885519, + 0.011440017260611057, + 0.0033161414321511984, + 0.003830170026049018, + 0.005116524174809456, + 0.0044409604743123055, + 0.00414758687838912, + 0.0035773313138633966, + 0.007677711080759764, + 0.004180398769676685, + 0.003807677188888192, + 0.0030561296734958887, + 0.005136186722666025, + 0.003852310124784708, + 0.0025959392078220844, + 0.003216475248336792, + 0.005417541600763798, + 0.0030202274210751057, + 0.00615402078256011, + 0.005353663582354784, + 0.005863995291292667, + 0.00516213895753026, + 0.0035970997996628284, + 0.005368640646338463, + 0.005442055873572826, + 0.003720949636772275, + 0.018177196383476257, + 0.012692316435277462, + 0.028498638421297073, + 0.020837990567088127, + 0.014586386270821095, + 0.014450098387897015, + 0.027154119685292244, + 0.042440857738256454, + 0.02353571355342865, + 0.011212441138923168, + 0.022775104269385338, + 0.013561592437326908, + 0.011187554337084293, + 0.012639857828617096, + 0.020194856449961662, + 0.01694575510919094, + 0.01627608947455883, + 0.0063287946395576, + 0.009233213029801846, + 0.008367905393242836, + 0.005248651374131441, + 0.005187050439417362, + 0.008592749014496803, + 0.004974942188709974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.020980216562747955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864844024181366 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.00864844024181366, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864844024181366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4030788242816925, + 0.060578182339668274, + 0.07358799874782562, + 0.05561784654855728, + 0.0874163955450058, + 0.04845410957932472, + 0.038368288427591324, + 0.11416301876306534, + 0.05481431633234024, + 0.0639210045337677 + ], + "credit_map": [ + 0.047090232372283936, + 0.03260966017842293, + 0.08381855487823486, + 0.039571743458509445, + 0.051068734377622604, + 0.037560123950242996, + 0.07700636237859726, + 0.03435341641306877, + 0.007651476189494133, + 0.007123017217963934, + 0.009127559140324593, + 0.007059059571474791, + 0.007401433773338795, + 0.0056536090560257435, + 0.0117497518658638, + 0.004812283907085657, + 0.009365984238684177, + 0.006912654265761375, + 0.008901858702301979, + 0.012126020155847073, + 0.0061303130351006985, + 0.005976730491966009, + 0.015772953629493713, + 0.008401486091315746, + 0.012440538965165615, + 0.007290690205991268, + 0.007173142395913601, + 0.006742157507687807, + 0.005435450468212366, + 0.005584569647908211, + 0.005442305468022823, + 0.0055089970119297504, + 0.010126341134309769, + 0.008743262849748135, + 0.018002111464738846, + 0.012696832418441772, + 0.00965364370495081, + 0.006210489198565483, + 0.014628473669290543, + 0.007355243433266878, + 0.005067199002951384, + 0.005795596167445183, + 0.007946344092488289, + 0.006241712719202042, + 0.004964438732713461, + 0.005630234256386757, + 0.007715383544564247, + 0.005093199200928211, + 0.006170247215777636, + 0.0054066115990281105, + 0.005123834125697613, + 0.004702839534729719, + 0.003156128339469433, + 0.005004269070923328, + 0.003481447696685791, + 0.00532291317358613, + 0.017855845391750336, + 0.011039872653782368, + 0.016011537984013557, + 0.013929595239460468, + 0.009353157132863998, + 0.009791146963834763, + 0.01574489288032055, + 0.020436976104974747, + 0.016089487820863724, + 0.004840081557631493, + 0.007998398505151272, + 0.0053624180145561695, + 0.004695766605436802, + 0.003939654678106308, + 0.00654596509411931, + 0.005342548247426748, + 0.01576942950487137, + 0.005985704250633717, + 0.009671091102063656, + 0.007823317311704159, + 0.005536903627216816, + 0.005790144205093384, + 0.0082781333476305, + 0.005066286772489548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.00864844024181366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072737164795399 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072737164795399, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072737164795399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1736202985048294, + 0.13625690340995789, + 0.06022277846932411, + 0.07697543501853943, + 0.06709398329257965, + 0.0336039774119854, + 0.10807861387729645, + 0.12005453556776047, + 0.11290373653173447, + 0.11118969321250916 + ], + "credit_map": [ + 0.01810949668288231, + 0.015360488556325436, + 0.04442647472023964, + 0.016792278736829758, + 0.016134170815348625, + 0.016118492931127548, + 0.029201477766036987, + 0.017477424815297127, + 0.01810801401734352, + 0.010215295478701591, + 0.037499140948057175, + 0.013949353247880936, + 0.012659729458391666, + 0.016864117234945297, + 0.013034604489803314, + 0.013926653191447258, + 0.009390432387590408, + 0.005266252439469099, + 0.008652107790112495, + 0.006389526184648275, + 0.006032898090779781, + 0.00615898659452796, + 0.01194668561220169, + 0.0063858916983008385, + 0.015585342422127724, + 0.010969910770654678, + 0.012006163597106934, + 0.007288485765457153, + 0.007093454711139202, + 0.0074751116335392, + 0.009299783036112785, + 0.007257184013724327, + 0.007410754449665546, + 0.005542626138776541, + 0.01295777503401041, + 0.008522630669176579, + 0.007849532179534435, + 0.005650574341416359, + 0.013962692581117153, + 0.0051974086090922356, + 0.00554475886747241, + 0.002960042329505086, + 0.004592335782945156, + 0.006769127212464809, + 0.002564753172919154, + 0.0023274885024875402, + 0.005746802780777216, + 0.0030986706260591745, + 0.032291535288095474, + 0.010907134972512722, + 0.015614059753715992, + 0.010546308010816574, + 0.007974392734467983, + 0.0076315514743328094, + 0.011399444192647934, + 0.011714193969964981, + 0.033908307552337646, + 0.012305689044296741, + 0.021512502804398537, + 0.011693461798131466, + 0.010160270147025585, + 0.011273818090558052, + 0.01100553572177887, + 0.008194957859814167, + 0.03253365308046341, + 0.010507786646485329, + 0.016434507444500923, + 0.009134449996054173, + 0.008274952881038189, + 0.0071162814274430275, + 0.014050723053514957, + 0.014851394109427929, + 0.028510065749287605, + 0.010310742072761059, + 0.01677333191037178, + 0.014293874613940716, + 0.009366113692522049, + 0.008752603083848953, + 0.013789905235171318, + 0.009393056854605675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011072737164795399 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772672317922115 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.009772672317922115, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772672317922115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39554354548454285, + 0.058278076350688934, + 0.02126535028219223, + 0.04516025260090828, + 0.07244129478931427, + 0.07383410632610321, + 0.10936827957630157, + 0.08045584708452225, + 0.07858234643936157, + 0.06507091969251633 + ], + "credit_map": [ + 0.04408218339085579, + 0.01529483962804079, + 0.12520387768745422, + 0.03254593908786774, + 0.026830006390810013, + 0.026820868253707886, + 0.08928963541984558, + 0.035476166754961014, + 0.009279001504182816, + 0.004639127291738987, + 0.013543661683797836, + 0.006477781105786562, + 0.005349719896912575, + 0.006790688261389732, + 0.006349543109536171, + 0.005848549772053957, + 0.0027473503723740578, + 0.0023089046590030193, + 0.0027738017961382866, + 0.0027001553680747747, + 0.0031999736092984676, + 0.002226511249318719, + 0.0032132852356880903, + 0.0020953661296516657, + 0.006136102136224508, + 0.0041679623536765575, + 0.00790184922516346, + 0.005852812435477972, + 0.005203757900744677, + 0.005040824878960848, + 0.007079480215907097, + 0.0037774555385112762, + 0.011242328211665154, + 0.006834594998508692, + 0.008113262243568897, + 0.011691159568727016, + 0.008624892681837082, + 0.006106298882514238, + 0.012296508997678757, + 0.007532243151217699, + 0.00922130886465311, + 0.005391272250562906, + 0.01576700620353222, + 0.00838380865752697, + 0.006436629220843315, + 0.0052004121243953705, + 0.017795346677303314, + 0.0056383125483989716, + 0.03708840161561966, + 0.011739673092961311, + 0.019292596727609634, + 0.009810819290578365, + 0.009042331948876381, + 0.008209963329136372, + 0.007775297854095697, + 0.006409183610230684, + 0.011586037464439869, + 0.0073356665670871735, + 0.01308722048997879, + 0.012604864314198494, + 0.006896705832332373, + 0.007326802704483271, + 0.01048317737877369, + 0.011135369539260864, + 0.019553570076823235, + 0.007041547447443008, + 0.010061479173600674, + 0.007520555984228849, + 0.006285202223807573, + 0.0060567609034478664, + 0.01234434638172388, + 0.009718869812786579, + 0.015592058189213276, + 0.006151712965220213, + 0.010744347237050533, + 0.007500707637518644, + 0.005182556342333555, + 0.005242590326815844, + 0.009484278969466686, + 0.0051726652309298515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.009772672317922115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304216057062149 + ], + "opposite_logprob_tokens": [ + -3.1680421829223633 + ], + "logprob_reward": -0.04304216057062149, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304216057062149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0343894399702549, + 0.016416877508163452, + 0.016240205615758896, + 0.020549438893795013, + 0.026033340021967888, + 0.01994103565812111, + 0.18788249790668488, + 0.23368369042873383, + 0.273508220911026, + 0.1713552176952362 + ], + "credit_map": [ + 0.004537394270300865, + 0.0023647556081414223, + 0.010149464942514896, + 0.0034249855671077967, + 0.0025677219964563847, + 0.002791106468066573, + 0.005410997197031975, + 0.003143016481772065, + 0.00263189896941185, + 0.0015989791136234999, + 0.002368014305830002, + 0.002025557914748788, + 0.0023569564800709486, + 0.0019493349827826023, + 0.0019814479164779186, + 0.0015046891057863832, + 0.0019297776743769646, + 0.0017330265836790204, + 0.0020069177262485027, + 0.002081133658066392, + 0.0016288727056235075, + 0.0017534906510263681, + 0.0033984726760536432, + 0.0017085130093619227, + 0.0020388688426464796, + 0.0032142801210284233, + 0.0030579552985727787, + 0.002822943963110447, + 0.0018291944870725274, + 0.0032550767064094543, + 0.0021013699006289244, + 0.0022297478280961514, + 0.0029561908449977636, + 0.002183800796046853, + 0.004403016995638609, + 0.003238328266888857, + 0.00274167419411242, + 0.0022648251615464687, + 0.0059009757824242115, + 0.002344529377296567, + 0.00386976869776845, + 0.0020433233585208654, + 0.0025862830225378275, + 0.0026750308461487293, + 0.001954471692442894, + 0.0018386599840596318, + 0.002789230551570654, + 0.002184267621487379, + 0.061091333627700806, + 0.01918891817331314, + 0.03011760301887989, + 0.018118198961019516, + 0.015602469444274902, + 0.014487069100141525, + 0.01654062233865261, + 0.012736272998154163, + 0.06886206567287445, + 0.023223871365189552, + 0.033126674592494965, + 0.0238187313079834, + 0.018251733854413033, + 0.0194883793592453, + 0.02143421769142151, + 0.025478027760982513, + 0.07666736841201782, + 0.02536468394100666, + 0.03196733444929123, + 0.024054840207099915, + 0.021217722445726395, + 0.020251234993338585, + 0.04048759117722511, + 0.033497441560029984, + 0.04409506544470787, + 0.01676693558692932, + 0.026976360008120537, + 0.020917236804962158, + 0.0141697246581316, + 0.013714724220335484, + 0.021806152537465096, + 0.01290901005268097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.04304216057062149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot shot from the ground, the world is reimagined as a child's drawing, brimming with playful innocence. At the center of the frame lies a forest of purple-lined, green-skinned bamboo. Each bamboo pole is not a perfect cylinder, but rather a slightly skewed, crayon-like line drawn with childish strokes. The color is unevenly filled, revealing a paper-like texture. Striking vertical purple stripes are meticulously inscribed on these apple-green poles, their edges tinged with the characteristic roughness of crayon. In the almost static flow of time, the bamboo forest sways with a dreamlike, almost imperceptible movement. The bamboo leaves, simplified into dark green paper-cut shapes, tremble subtly with the swaying of the poles. Most strikingly, with each subtle bend of the poles, the purple stripes appear to slowly expand and contract, creating a sense of life, akin to \"breathing.\" The background is a flat, soft ochre earth and azure sky, dotted with a few fluffy white clouds like cotton candy. In the corner, a huge yellow sun with a smiling face casts warm and even light, injecting infinite tenderness and reverie into this quiet, innocent fairy tale scene.", + "seed": 21423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_12/index_3408_seed_21423_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03935739668086171, + "vlm_logprob": -0.03935739668086171, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03935739668086171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030244745314121246 + ], + "opposite_logprob_tokens": [ + -3.530244827270508 + ], + "logprob_reward": -0.030244745314121246, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030244745314121246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12028820067644119, + 0.08175404369831085, + 0.03775568678975105, + 0.1765763759613037, + 0.13110995292663574, + 0.17633944749832153, + 0.04438761621713638, + 0.09398793429136276, + 0.1045866459608078, + 0.03321416676044464 + ], + "credit_map": [ + 0.009980562143027782, + 0.02804834209382534, + 0.01821163482964039, + 0.016009438782930374, + 0.011129808612167835, + 0.013328418135643005, + 0.010958909057080746, + 0.012621085159480572, + 0.007220205385237932, + 0.016386179253458977, + 0.015879856422543526, + 0.010014784522354603, + 0.008243048563599586, + 0.009833769872784615, + 0.0072119818069040775, + 0.00696420855820179, + 0.0029339916072785854, + 0.00670937355607748, + 0.0050238389521837234, + 0.007243702653795481, + 0.003386018332093954, + 0.0037342016585171223, + 0.0042122360318899155, + 0.004512321669608355, + 0.010408466681838036, + 0.04811298847198486, + 0.012788686901330948, + 0.061565157026052475, + 0.010418150573968887, + 0.007589065004140139, + 0.0182801503688097, + 0.007413689978420734, + 0.010381806641817093, + 0.01961676962673664, + 0.012416670098900795, + 0.03403870388865471, + 0.016203327104449272, + 0.010085444897413254, + 0.01961320824921131, + 0.008754014037549496, + 0.008654290810227394, + 0.04447496682405472, + 0.021427245810627937, + 0.06350941210985184, + 0.01014381181448698, + 0.00698505574837327, + 0.014060971327126026, + 0.007083673495799303, + 0.003087886841967702, + 0.0065818228758871555, + 0.006020333152264357, + 0.008023262955248356, + 0.004044806119054556, + 0.004925363231450319, + 0.004487191326916218, + 0.007216946687549353, + 0.0073034875094890594, + 0.01725492998957634, + 0.010726683773100376, + 0.011637052521109581, + 0.019920215010643005, + 0.00941469520330429, + 0.009467674419283867, + 0.008263186551630497, + 0.010740324854850769, + 0.011934779584407806, + 0.006592107005417347, + 0.01431784313172102, + 0.021760093048214912, + 0.014376761391758919, + 0.00942760705947876, + 0.015437123365700245, + 0.003239275189116597, + 0.006095688324421644, + 0.0036048355977982283, + 0.006108960136771202, + 0.004077292047441006, + 0.0029997709207236767, + 0.004158369265496731, + 0.002929973416030407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.030244745314121246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307607561349869 + ], + "opposite_logprob_tokens": [ + -3.1680760383605957 + ], + "logprob_reward": -0.04307607561349869, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307607561349869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06359563022851944, + 0.06036145240068436, + 0.03586239740252495, + 0.20150235295295715, + 0.05586978793144226, + 0.4069617688655853, + 0.08571846038103104, + 0.030964558944106102, + 0.03864177688956261, + 0.020521804690361023 + ], + "credit_map": [ + 0.005843362770974636, + 0.007637538481503725, + 0.013264214619994164, + 0.008856077678501606, + 0.008344613946974277, + 0.005862231831997633, + 0.0069959587417542934, + 0.0067916312254965305, + 0.005732376128435135, + 0.007138555869460106, + 0.011094818823039532, + 0.009236534126102924, + 0.005942170973867178, + 0.007936340756714344, + 0.007028939202427864, + 0.006251719780266285, + 0.003371919272467494, + 0.009641793556511402, + 0.004226479679346085, + 0.007103984709829092, + 0.0033607380464673042, + 0.002730723936110735, + 0.002926752669736743, + 0.002500007161870599, + 0.014725111424922943, + 0.05824192240834236, + 0.011894450522959232, + 0.06556913256645203, + 0.015763521194458008, + 0.009984434582293034, + 0.01771712675690651, + 0.007606674451380968, + 0.004167557694017887, + 0.009953497909009457, + 0.005248075816780329, + 0.014486641623079777, + 0.007820029743015766, + 0.0036455767694860697, + 0.00646542152389884, + 0.004082989878952503, + 0.01801282726228237, + 0.12914416193962097, + 0.021617820486426353, + 0.1617027372121811, + 0.01360342651605606, + 0.012795446440577507, + 0.041116129606962204, + 0.008969222195446491, + 0.0046670399606227875, + 0.023259077221155167, + 0.006239980924874544, + 0.026912296190857887, + 0.0072774565778672695, + 0.005290891509503126, + 0.008099205791950226, + 0.003972507547587156, + 0.002996730851009488, + 0.004623109940439463, + 0.004692464601248503, + 0.006424521561712027, + 0.002560628345236182, + 0.0027929760981351137, + 0.003183158114552498, + 0.0036909671034663916, + 0.0033767945133149624, + 0.004645784851163626, + 0.0024596210569143295, + 0.004667812492698431, + 0.006880476605147123, + 0.006259635556489229, + 0.004153097979724407, + 0.006198553368449211, + 0.0017073978669941425, + 0.003273025155067444, + 0.001462622545659542, + 0.0037700145039707422, + 0.0034019642043858767, + 0.0027693030424416065, + 0.0018969913944602013, + 0.002240485744550824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.04307607561349869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944667384028435 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944667384028435, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944667384028435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10645116120576859, + 0.05427979677915573, + 0.09323833137750626, + 0.4107092320919037, + 0.09979867935180664, + 0.05848705396056175, + 0.0932566225528717, + 0.029818587005138397, + 0.033761296421289444, + 0.02019927278161049 + ], + "credit_map": [ + 0.00930389016866684, + 0.03440984711050987, + 0.014491778798401356, + 0.015599464066326618, + 0.009391852654516697, + 0.009247178211808205, + 0.008326034061610699, + 0.005681112874299288, + 0.005591325927525759, + 0.015553025528788567, + 0.006440719589591026, + 0.008495147339999676, + 0.00511125335469842, + 0.004946928471326828, + 0.003864874364808202, + 0.004276520572602749, + 0.008939068764448166, + 0.019035717472434044, + 0.018551796674728394, + 0.014664054848253727, + 0.006690043490380049, + 0.005481329280883074, + 0.010548682883381844, + 0.009327633306384087, + 0.02469995804131031, + 0.12394294887781143, + 0.015589158982038498, + 0.16002431511878967, + 0.015443063341081142, + 0.012917133048176765, + 0.046729955822229385, + 0.011362692341208458, + 0.007041714154183865, + 0.01484488695859909, + 0.009541783481836319, + 0.018064627423882484, + 0.02203645557165146, + 0.007852895185351372, + 0.011061201803386211, + 0.009355118498206139, + 0.004862009081989527, + 0.0065236822701990604, + 0.013616240583360195, + 0.010564112104475498, + 0.009144607000052929, + 0.0035628376062959433, + 0.005287322215735912, + 0.004926244728267193, + 0.007291669491678476, + 0.01642812229692936, + 0.012393221259117126, + 0.013995572924613953, + 0.016826918348670006, + 0.010084511712193489, + 0.007372716907411814, + 0.008863890543580055, + 0.002312858821824193, + 0.006022770423442125, + 0.003505706787109375, + 0.0045893751084804535, + 0.004420574754476547, + 0.0028113333974033594, + 0.0030940480064600706, + 0.003061919705942273, + 0.0030902547296136618, + 0.004603976383805275, + 0.0027310538571327925, + 0.004139965865761042, + 0.007196841295808554, + 0.004526779521256685, + 0.003342514391988516, + 0.004129910375922918, + 0.0019417229341343045, + 0.0033777528442442417, + 0.003044211771339178, + 0.0028320106212049723, + 0.002532604616135359, + 0.0026369327679276466, + 0.0016836029244586825, + 0.0021504336036741734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005944667384028435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631743300706148 + ], + "opposite_logprob_tokens": [ + -5.379631519317627 + ], + "logprob_reward": -0.004631743300706148, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631743300706148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13196870684623718, + 0.09702219814062119, + 0.34905970096588135, + 0.10466201603412628, + 0.04457549750804901, + 0.15582969784736633, + 0.03483074903488159, + 0.03004913590848446, + 0.03648661449551582, + 0.015515727922320366 + ], + "credit_map": [ + 0.009752818383276463, + 0.023784294724464417, + 0.02003556489944458, + 0.013522399589419365, + 0.015301277860999107, + 0.015381842851638794, + 0.01946134865283966, + 0.014729158952832222, + 0.008361328393220901, + 0.013249697163701057, + 0.011702227406203747, + 0.014910107478499413, + 0.013072636909782887, + 0.010228751227259636, + 0.015244993381202221, + 0.010252450592815876, + 0.0357077494263649, + 0.07347632944583893, + 0.077468141913414, + 0.054785579442977905, + 0.022590380162000656, + 0.018622994422912598, + 0.03763042017817497, + 0.02877807803452015, + 0.006212624721229076, + 0.03011764958500862, + 0.007623924873769283, + 0.03527679294347763, + 0.0053542000241577625, + 0.005374147091060877, + 0.009759102948009968, + 0.004943562671542168, + 0.0035349309910088778, + 0.007053579203784466, + 0.006426744628697634, + 0.008591393940150738, + 0.00627807155251503, + 0.003405499504879117, + 0.0055340975522994995, + 0.003751178039237857, + 0.007135600782930851, + 0.04043552652001381, + 0.01542703527957201, + 0.061050936579704285, + 0.007862954400479794, + 0.005382858216762543, + 0.014602513052523136, + 0.003932263236492872, + 0.0019296284299343824, + 0.0065850610844790936, + 0.0027139291632920504, + 0.010393896140158176, + 0.0043838354758918285, + 0.002712822053581476, + 0.0037337960675358772, + 0.0023777782917022705, + 0.002208283171057701, + 0.004655453842133284, + 0.0042841630056500435, + 0.005471592769026756, + 0.004573422484099865, + 0.0024060201831161976, + 0.0036216771695762873, + 0.0028285211883485317, + 0.002842349698767066, + 0.007236073259264231, + 0.0028244603890925646, + 0.011835680343210697, + 0.0027614221908152103, + 0.0023254279512912035, + 0.004387913271784782, + 0.0022732808720320463, + 0.0015197148313745856, + 0.002706131897866726, + 0.0021067929919809103, + 0.0029443122912198305, + 0.0016728714108467102, + 0.0016903480282053351, + 0.0017585827736184, + 0.0011169714853167534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.004631743300706148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene depicted as a child's drawing with a paper-like texture and crayon strokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo poles green with unevenly filled purple stripes and rough edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo leaves simplified into dark green paper-cut shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of flat ochre earth and an azure sky with fluffy white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a huge yellow sun with a smiling face in the corner casting warm light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11288975179195404 + ], + "opposite_logprob_tokens": [ + -2.2378897666931152 + ], + "logprob_reward": -0.11288975179195404, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11288975179195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12917189300060272, + 0.025084156543016434, + 0.08112454414367676, + 0.2229003608226776, + 0.07312744110822678, + 0.28556132316589355, + 0.06893537193536758, + 0.04171183705329895, + 0.021592337638139725, + 0.0507906898856163 + ], + "credit_map": [ + 0.00994475930929184, + 0.03503446280956268, + 0.020044587552547455, + 0.017925165593624115, + 0.010478483512997627, + 0.011666982434689999, + 0.013313306495547295, + 0.010764149948954582, + 0.0020555523224174976, + 0.003472547745332122, + 0.0029974619392305613, + 0.003515451680868864, + 0.0031298778485506773, + 0.0032247405033558607, + 0.003769181901589036, + 0.0029193430673331022, + 0.007512343116104603, + 0.014821559190750122, + 0.017421532422304153, + 0.013714415021240711, + 0.005852696020156145, + 0.005022774916142225, + 0.008888105861842632, + 0.007891122251749039, + 0.01352491695433855, + 0.0669533908367157, + 0.009183334186673164, + 0.08539210259914398, + 0.008875248022377491, + 0.007332898676395416, + 0.025211701169610023, + 0.00642678327858448, + 0.004953359719365835, + 0.008796988055109978, + 0.012986053712666035, + 0.014185100793838501, + 0.010266506113111973, + 0.005053486209362745, + 0.009913371875882149, + 0.0069725695066154, + 0.01260324940085411, + 0.08705465495586395, + 0.02008960396051407, + 0.11412636190652847, + 0.009503433480858803, + 0.008792650885879993, + 0.02738807164132595, + 0.006003305781632662, + 0.003392418846487999, + 0.018812278285622597, + 0.004351553972810507, + 0.02400083653628826, + 0.004485841374844313, + 0.0032078369986265898, + 0.0071113621816039085, + 0.0035732451360672712, + 0.004789932630956173, + 0.007978246547281742, + 0.007671802304685116, + 0.005820040125399828, + 0.004790363367646933, + 0.0028056500013917685, + 0.0031799080315977335, + 0.004675897303968668, + 0.0021281400695443153, + 0.0033855901565402746, + 0.0025549137499183416, + 0.0032487697899341583, + 0.003668260993435979, + 0.0021015277598053217, + 0.0022375728003680706, + 0.00226756208576262, + 0.005774971097707748, + 0.00861109048128128, + 0.0055901468731462955, + 0.00634877709671855, + 0.007290631998330355, + 0.005929312203079462, + 0.0065467096865177155, + 0.004699050448834896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.11288975179195404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_tensor([5658, 5658])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_tensor([5658, 5658])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e46ed30bfc9302660c8abadb44a98f1ede41ee8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_tensor([5658, 5658])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking, high-contrast black-and-white overhead shot, the camera held still, focused vertically on the center of a dark cast-iron pot below. Inside, a boiling, nearly pure white broth formed a violent vortex. A whole fish—head and tail intact—tumbled uncontrollably in this turmoil. Illuminated by the harsh overhead light, the fish's slippery scales reflected dazzling white highlights, contrasting starkly with its deep black skin. Its body twisted and writhed violently with the onslaught of the scalding liquid, sometimes flung to the surface, revealing its pale underbelly, and sometimes completely engulfed by the churning bubbles and opaque broth. The fish's hollow eyes were deep black spots, its mouth slightly open, as if in a silent cry. Each tumble stirred the broth, sending its tail fin slapping violently against the surface, sending up countless bright white droplets that traced brief, sharp arcs in the air before falling back onto the inky black pot wall and the bubbling white broth. The entire picture is full of primitive and wild dynamic tension, creating an intense, tense and almost violent culinary atmosphere.", + "seed": 4761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4761_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04388524126261473, + "vlm_logprob": -0.04388524126261473, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04388524126261473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03578752279281616 + ], + "opposite_logprob_tokens": [ + -3.785787582397461 + ], + "logprob_reward": -0.03578752279281616, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03578752279281616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2590661644935608, + 0.09379968792200089, + 0.05052759498357773, + 0.05543304607272148, + 0.18685394525527954, + 0.0826745554804802, + 0.08174760639667511, + 0.03718779608607292, + 0.04486199468374252, + 0.10784757137298584 + ], + "credit_map": [ + 0.019946396350860596, + 0.03454585373401642, + 0.0423860140144825, + 0.04263148456811905, + 0.03038245067000389, + 0.025138836354017258, + 0.03355430066585541, + 0.03048083558678627, + 0.00469324691221118, + 0.010174294002354145, + 0.010006793774664402, + 0.027990253642201424, + 0.011706892400979996, + 0.0049602496437728405, + 0.015210174024105072, + 0.009057799354195595, + 0.005010964348912239, + 0.00513832550495863, + 0.0052958386950194836, + 0.008457674644887447, + 0.004700671415776014, + 0.004106651060283184, + 0.01070865336805582, + 0.007108826655894518, + 0.004167937207967043, + 0.006384182721376419, + 0.008046336472034454, + 0.010141356848180294, + 0.0050992644391953945, + 0.004162917844951153, + 0.01256298553198576, + 0.0048680733889341354, + 0.007642153650522232, + 0.03153550624847412, + 0.005887419451028109, + 0.01234804093837738, + 0.009178821928799152, + 0.035981398075819016, + 0.07536712288856506, + 0.008913502097129822, + 0.0036085788160562515, + 0.008530398830771446, + 0.007557328790426254, + 0.012131759896874428, + 0.0094170356169343, + 0.007897188887000084, + 0.02451128326356411, + 0.009020992554724216, + 0.006093637552112341, + 0.014544653706252575, + 0.012829060666263103, + 0.008638840168714523, + 0.010012499988079071, + 0.006881481967866421, + 0.010887240059673786, + 0.011860202066600323, + 0.002995380200445652, + 0.003770738374441862, + 0.004270127508789301, + 0.0033155353739857674, + 0.007668974343687296, + 0.003993637394160032, + 0.005417061969637871, + 0.005756341852247715, + 0.0044091022573411465, + 0.003926279954612255, + 0.005384877324104309, + 0.005780056118965149, + 0.004200613591820002, + 0.006426979321986437, + 0.00905133318156004, + 0.005682758055627346, + 0.00812641903758049, + 0.02273300476372242, + 0.006871060933917761, + 0.010308954864740372, + 0.010433864779770374, + 0.018976358696818352, + 0.020157132297754288, + 0.01024079043418169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03578752279281616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296333372592926 + ], + "opposite_logprob_tokens": [ + -2.2379634380340576 + ], + "logprob_reward": -0.11296333372592926, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296333372592926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1984257698059082, + 0.07717841118574142, + 0.09685724973678589, + 0.05087845027446747, + 0.2006545513868332, + 0.043590690940618515, + 0.13676851987838745, + 0.06025083735585213, + 0.05797398090362549, + 0.07742144167423248 + ], + "credit_map": [ + 0.01397181861102581, + 0.025310451164841652, + 0.02637510560452938, + 0.0409211702644825, + 0.017638253048062325, + 0.01779533177614212, + 0.027961090207099915, + 0.028452537953853607, + 0.0072969552129507065, + 0.007242672611027956, + 0.011141983792185783, + 0.011533444747328758, + 0.011773396283388138, + 0.007817606441676617, + 0.009654232300817966, + 0.010718118399381638, + 0.008889825083315372, + 0.010309633798897266, + 0.012970713898539543, + 0.0156334787607193, + 0.009193914011120796, + 0.009041658602654934, + 0.022134605795145035, + 0.008683416992425919, + 0.0028883370105177164, + 0.00686541898176074, + 0.0072302669286727905, + 0.008399846032261848, + 0.00544339744374156, + 0.002810052363201976, + 0.011924435384571552, + 0.0053166975267231464, + 0.0084672337397933, + 0.041104793548583984, + 0.006360176485031843, + 0.01240365207195282, + 0.012705609202384949, + 0.031040748581290245, + 0.0751715674996376, + 0.013400787487626076, + 0.003038264811038971, + 0.004819149151444435, + 0.0037308442406356335, + 0.008251532912254333, + 0.004762046970427036, + 0.007062898017466068, + 0.006209235172718763, + 0.0057167233899235725, + 0.012998255901038647, + 0.014251239597797394, + 0.019407078623771667, + 0.014595038257539272, + 0.014862542040646076, + 0.016258733347058296, + 0.019739815965294838, + 0.024655824527144432, + 0.007203403394669294, + 0.0066455816850066185, + 0.005715020466595888, + 0.004101424477994442, + 0.009364617057144642, + 0.006421463098376989, + 0.008148208260536194, + 0.012651119381189346, + 0.003740705316886306, + 0.006157266907393932, + 0.005816549528390169, + 0.009525437839329243, + 0.005843145307153463, + 0.0077045392245054245, + 0.013276640325784683, + 0.005909698083996773, + 0.00628423597663641, + 0.009371497668325901, + 0.006864568684250116, + 0.010174489580094814, + 0.0064019919373095036, + 0.01253130566328764, + 0.014468926005065441, + 0.011324426159262657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.11296333372592926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803181067109108 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.029803181067109108, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803181067109108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08459751307964325, + 0.04359396919608116, + 0.06219334527850151, + 0.07343023270368576, + 0.2593517303466797, + 0.20073087513446808, + 0.049032535403966904, + 0.06478497385978699, + 0.07577764242887497, + 0.08650717884302139 + ], + "credit_map": [ + 0.006051763892173767, + 0.0065638902597129345, + 0.01385957095772028, + 0.018681399524211884, + 0.006947699934244156, + 0.009926171973347664, + 0.012602186761796474, + 0.009964828379452229, + 0.0032029799185693264, + 0.004650391638278961, + 0.004473454784601927, + 0.010815274901688099, + 0.004878575913608074, + 0.003094523446634412, + 0.006938156671822071, + 0.005540609825402498, + 0.005011022090911865, + 0.009079339914023876, + 0.007999645546078682, + 0.00935619231313467, + 0.004890359938144684, + 0.004101224709302187, + 0.014898947440087795, + 0.0068566142581403255, + 0.0041459836065769196, + 0.011116943322122097, + 0.010553187690675259, + 0.010957453399896622, + 0.008417409844696522, + 0.005735330283641815, + 0.014678172767162323, + 0.007825751788914204, + 0.014901851303875446, + 0.013230803422629833, + 0.011852970346808434, + 0.018903415650129318, + 0.012571454979479313, + 0.08632632344961166, + 0.08556631952524185, + 0.015998607501387596, + 0.014241310767829418, + 0.02228650450706482, + 0.018313763663172722, + 0.04533860459923744, + 0.021730398759245872, + 0.020923329517245293, + 0.039009224623441696, + 0.018887732177972794, + 0.004984520375728607, + 0.005526162218302488, + 0.007298762444406748, + 0.005259308498352766, + 0.0051910108886659145, + 0.005316324066370726, + 0.007255288772284985, + 0.008201157674193382, + 0.007937840186059475, + 0.005733646918088198, + 0.0071142930537462234, + 0.00802033580839634, + 0.008058742620050907, + 0.007886205799877644, + 0.0071381693705916405, + 0.012895737774670124, + 0.005641996394842863, + 0.00613967701792717, + 0.008488522842526436, + 0.011711252853274345, + 0.0077222976833581924, + 0.012267070822417736, + 0.01546518411487341, + 0.008341646753251553, + 0.007704799994826317, + 0.013945733197033405, + 0.0063724517822265625, + 0.011051210574805737, + 0.009108149446547031, + 0.01634921506047249, + 0.010806499049067497, + 0.011169113218784332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.029803181067109108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10739785432815552, + 0.03978833928704262, + 0.054749395698308945, + 0.026309220120310783, + 0.30272647738456726, + 0.2932330071926117, + 0.034993499517440796, + 0.03860366716980934, + 0.04068297892808914, + 0.06151556596159935 + ], + "credit_map": [ + 0.007054776884615421, + 0.010659204795956612, + 0.017109926789999008, + 0.01861848309636116, + 0.012308062054216862, + 0.011343506164848804, + 0.014081668108701706, + 0.01622222177684307, + 0.003919737879186869, + 0.004512906540185213, + 0.0033090277574956417, + 0.009222093969583511, + 0.005743222311139107, + 0.0036250848788768053, + 0.005430799443274736, + 0.004025466274470091, + 0.003895573318004608, + 0.005310556851327419, + 0.005197281017899513, + 0.010322866030037403, + 0.0038806884549558163, + 0.004504759795963764, + 0.015829643234610558, + 0.005808028392493725, + 0.0016876746667549014, + 0.0035418588668107986, + 0.004560037981718779, + 0.004034537822008133, + 0.0023053192999213934, + 0.001507000532001257, + 0.0061829532496631145, + 0.0024898385163396597, + 0.014044394716620445, + 0.027384815737605095, + 0.011499680578708649, + 0.018411938101053238, + 0.014549916610121727, + 0.08115827292203903, + 0.11954313516616821, + 0.016134338453412056, + 0.008966835215687752, + 0.022445185109972954, + 0.02275189757347107, + 0.034620124846696854, + 0.029058678075671196, + 0.04274464398622513, + 0.10962057113647461, + 0.023025082424283028, + 0.003123678034171462, + 0.004307915456593037, + 0.005063624586910009, + 0.004162658471614122, + 0.0036014686338603497, + 0.0037584330420941114, + 0.005229168105870485, + 0.0057465508580207825, + 0.0030546982306987047, + 0.004406472202390432, + 0.004353872500360012, + 0.00481814332306385, + 0.007485930807888508, + 0.00445524649694562, + 0.004553907550871372, + 0.005475393962115049, + 0.003392179496586323, + 0.004172327928245068, + 0.0038294291589409113, + 0.00625188834965229, + 0.0043061114847660065, + 0.006983997765928507, + 0.00712027819827199, + 0.00462676864117384, + 0.004143333062529564, + 0.008547271601855755, + 0.005935323424637318, + 0.009105668403208256, + 0.006862124893814325, + 0.008556762710213661, + 0.00923413597047329, + 0.009130950085818768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789527878165245 + ], + "opposite_logprob_tokens": [ + -3.529789447784424 + ], + "logprob_reward": -0.029789527878165245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789527878165245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08342118561267853, + 0.06994091719388962, + 0.09340856224298477, + 0.040314652025699615, + 0.1730678230524063, + 0.2746536135673523, + 0.023310720920562744, + 0.07941416651010513, + 0.059232860803604126, + 0.10323550552129745 + ], + "credit_map": [ + 0.007671101950109005, + 0.010670699179172516, + 0.01178025733679533, + 0.012918488122522831, + 0.007927628234028816, + 0.009691338054835796, + 0.010431930422782898, + 0.012329746969044209, + 0.004839810077100992, + 0.006754460744559765, + 0.010412507690489292, + 0.013029184192419052, + 0.010659324936568737, + 0.005962710361927748, + 0.00880486611276865, + 0.009478055872023106, + 0.008262231945991516, + 0.007568058557808399, + 0.013604489155113697, + 0.01689665950834751, + 0.00811544619500637, + 0.010906049981713295, + 0.01862873136997223, + 0.009426896460354328, + 0.0031394041143357754, + 0.006864708848297596, + 0.00641949288547039, + 0.006702017039060593, + 0.0037377967964857817, + 0.002686080988496542, + 0.007229076698422432, + 0.0035360727924853563, + 0.007479764521121979, + 0.03329053148627281, + 0.004654877819120884, + 0.007477269973605871, + 0.01141590066254139, + 0.028283216059207916, + 0.06951415538787842, + 0.01095210388302803, + 0.014041250571608543, + 0.021193893626332283, + 0.026215163990855217, + 0.029439741745591164, + 0.028114937245845795, + 0.0476679690182209, + 0.08213099092245102, + 0.025849686935544014, + 0.0016059757908806205, + 0.003229751018807292, + 0.0030957693234086037, + 0.0028657326474785805, + 0.003173430683091283, + 0.0024335812777280807, + 0.003163177752867341, + 0.003743302309885621, + 0.007109303958714008, + 0.010407119058072567, + 0.009291596710681915, + 0.009706346318125725, + 0.009993334300816059, + 0.008560026064515114, + 0.01312045194208622, + 0.011225986294448376, + 0.00439359899610281, + 0.0050573269836604595, + 0.006722490303218365, + 0.008240633644163609, + 0.006289167795330286, + 0.010086060501635075, + 0.011613912880420685, + 0.006829667370766401, + 0.009707823395729065, + 0.015353883616626263, + 0.008401378989219666, + 0.013723960146307945, + 0.012406846508383751, + 0.019555723294615746, + 0.01248499657958746, + 0.011600885540246964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.029789527878165245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, high-contrast black-and-white overhead shot, the camera held still, focused vertically on the center of a dark cast-iron pot below. Inside, a boiling, nearly pure white broth formed a violent vortex. A whole fish—head and tail intact—tumbled uncontrollably in this turmoil. Illuminated by the harsh overhead light, the fish's slippery scales reflected dazzling white highlights, contrasting starkly with its deep black skin. Its body twisted and writhed violently with the onslaught of the scalding liquid, sometimes flung to the surface, revealing its pale underbelly, and sometimes completely engulfed by the churning bubbles and opaque broth. The fish's hollow eyes were deep black spots, its mouth slightly open, as if in a silent cry. Each tumble stirred the broth, sending its tail fin slapping violently against the surface, sending up countless bright white droplets that traced brief, sharp arcs in the air before falling back onto the inky black pot wall and the bubbling white broth. The entire picture is full of primitive and wild dynamic tension, creating an intense, tense and almost violent culinary atmosphere.", + "seed": 4762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4762_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02811216190457344, + "vlm_logprob": -0.02811216190457344, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02811216190457344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016540072858333588 + ], + "opposite_logprob_tokens": [ + -5.016540050506592 + ], + "logprob_reward": -0.016540072858333588, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016540072858333588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15142636001110077, + 0.07267103344202042, + 0.07615206390619278, + 0.054842546582221985, + 0.03940488025546074, + 0.08569968491792679, + 0.04440649226307869, + 0.03376523032784462, + 0.2075708508491516, + 0.2340608686208725 + ], + "credit_map": [ + 0.01652652397751808, + 0.02330746315419674, + 0.01835802011191845, + 0.024240415543317795, + 0.009117177687585354, + 0.012997858226299286, + 0.025931697338819504, + 0.02094719558954239, + 0.013578437268733978, + 0.009598219767212868, + 0.008311604149639606, + 0.009137030690908432, + 0.008856593631207943, + 0.008215043693780899, + 0.008635470643639565, + 0.006338630802929401, + 0.005884696263819933, + 0.0068221306428313255, + 0.0160691998898983, + 0.01167092565447092, + 0.009430370293557644, + 0.010461244732141495, + 0.00952623039484024, + 0.006287256721407175, + 0.0049407887272536755, + 0.0037143093068152666, + 0.008255649358034134, + 0.009826533496379852, + 0.005326224025338888, + 0.0052015031687915325, + 0.008011563681066036, + 0.009565968997776508, + 0.003211397211998701, + 0.005925068166106939, + 0.00545385479927063, + 0.0048278155736625195, + 0.006351029500365257, + 0.004750377964228392, + 0.004540928173810244, + 0.004344405606389046, + 0.007503170054405928, + 0.012878023087978363, + 0.012069611810147762, + 0.009764538146555424, + 0.017736302688717842, + 0.00957693625241518, + 0.007501659914851189, + 0.00866943970322609, + 0.0038800621405243874, + 0.005261396057903767, + 0.006618305575102568, + 0.006240516435354948, + 0.004211187828332186, + 0.005149354226887226, + 0.007174529135227203, + 0.005871139466762543, + 0.0024256291799247265, + 0.00340057210996747, + 0.004753263667225838, + 0.004669137764722109, + 0.00370426825247705, + 0.00394262233749032, + 0.005385644733905792, + 0.005484091583639383, + 0.013813422061502934, + 0.022803258150815964, + 0.022876640781760216, + 0.02698507532477379, + 0.05546467378735542, + 0.01967141218483448, + 0.024615753442049026, + 0.021340586245059967, + 0.007938100025057793, + 0.013762097805738449, + 0.010974065400660038, + 0.08528023213148117, + 0.007127515040338039, + 0.01194016169756651, + 0.02683531679213047, + 0.07020338624715805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016540072858333588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207158789038658 + ], + "opposite_logprob_tokens": [ + -2.8120715618133545 + ], + "logprob_reward": -0.06207158789038658, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207158789038658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09210003912448883, + 0.05532708019018173, + 0.04853741452097893, + 0.06283189356327057, + 0.048364218324422836, + 0.07202381640672684, + 0.03897266834974289, + 0.11303819715976715, + 0.04299081489443779, + 0.42581382393836975 + ], + "credit_map": [ + 0.011714757420122623, + 0.01344213541597128, + 0.014165082015097141, + 0.010106854140758514, + 0.008277633227407932, + 0.009506743401288986, + 0.013164184056222439, + 0.011722652241587639, + 0.007955849170684814, + 0.0076648881658911705, + 0.006165704689919949, + 0.007704556453973055, + 0.007725848350673914, + 0.0050880941562354565, + 0.007772407028824091, + 0.005249734967947006, + 0.004532052204012871, + 0.005335628055036068, + 0.009551198221743107, + 0.007251998409628868, + 0.005053916946053505, + 0.005080237518996, + 0.007202095817774534, + 0.004530290607362986, + 0.006613684818148613, + 0.00539070088416338, + 0.006880282424390316, + 0.00959178525954485, + 0.006862164940685034, + 0.006632869131863117, + 0.009867684915661812, + 0.010992727242410183, + 0.0037528707180172205, + 0.006184566766023636, + 0.007136969361454248, + 0.008436933159828186, + 0.006165793631225824, + 0.004900271072983742, + 0.0057898941449820995, + 0.005996923428028822, + 0.007228244096040726, + 0.009966952726244926, + 0.009370190091431141, + 0.008063653483986855, + 0.016219137236475945, + 0.007122152950614691, + 0.008228634484112263, + 0.005824852269142866, + 0.002804641844704747, + 0.003425013739615679, + 0.00514715351164341, + 0.005927102640271187, + 0.004246805794537067, + 0.0036315815523266792, + 0.00969564076513052, + 0.004094729665666819, + 0.009301526471972466, + 0.01113919634371996, + 0.022395029664039612, + 0.018108349293470383, + 0.009021470323204994, + 0.008162922225892544, + 0.022051339969038963, + 0.0128583749756217, + 0.0019361961167305708, + 0.002679156605154276, + 0.0029723416082561016, + 0.013561667874455452, + 0.0031880659516900778, + 0.0034984112717211246, + 0.0040591866709291935, + 0.011095794849097729, + 0.014023132622241974, + 0.025850841775536537, + 0.015094282105565071, + 0.15902531147003174, + 0.010481084696948528, + 0.0194759052246809, + 0.0624656118452549, + 0.11939769983291626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.06207158789038658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070479527115822 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070479527115822, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070479527115822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07405496388673782, + 0.030495535582304, + 0.09153815358877182, + 0.021996449679136276, + 0.035282254219055176, + 0.05654510483145714, + 0.044548239558935165, + 0.09581926465034485, + 0.08418628573417664, + 0.4655337333679199 + ], + "credit_map": [ + 0.009092822670936584, + 0.011381353251636028, + 0.010107708163559437, + 0.006841282360255718, + 0.00520701427012682, + 0.0061693997122347355, + 0.015127125196158886, + 0.010128260590136051, + 0.0040272981859743595, + 0.00485765328630805, + 0.00301508791744709, + 0.003929587081074715, + 0.004581657238304615, + 0.0028163250535726547, + 0.0038530349265784025, + 0.003414892591536045, + 0.009215672500431538, + 0.008799922652542591, + 0.011084356345236301, + 0.011734392493963242, + 0.019054405391216278, + 0.013776542618870735, + 0.00944509170949459, + 0.008427761495113373, + 0.0023202132433652878, + 0.0021993666887283325, + 0.0024613491259515285, + 0.0036678710021078587, + 0.002266440773382783, + 0.002653984585776925, + 0.0032497525680810213, + 0.0031774735543876886, + 0.0024395540822297335, + 0.00418486213311553, + 0.004087760578840971, + 0.006786614656448364, + 0.0039725108072161674, + 0.004287898074835539, + 0.0038414911832660437, + 0.005681563634425402, + 0.0035697687417268753, + 0.009000707417726517, + 0.005852884612977505, + 0.006497287657111883, + 0.012871808372437954, + 0.003849505679681897, + 0.010005619376897812, + 0.004897521808743477, + 0.0030809773597866297, + 0.003994661848992109, + 0.004382101818919182, + 0.006957218982279301, + 0.006557706743478775, + 0.004153869580477476, + 0.01140596903860569, + 0.004015732556581497, + 0.00894018542021513, + 0.008898216299712658, + 0.019801247864961624, + 0.01518848817795515, + 0.00552242761477828, + 0.008516411297023296, + 0.0182870551943779, + 0.010665238834917545, + 0.004124720115214586, + 0.004931359551846981, + 0.005502595566213131, + 0.026111500337719917, + 0.010058419778943062, + 0.005589914508163929, + 0.010321434587240219, + 0.017546340823173523, + 0.015148421749472618, + 0.023490654304623604, + 0.019387708976864815, + 0.14775069057941437, + 0.01472163014113903, + 0.025113020092248917, + 0.04333781450986862, + 0.17658379673957825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.016070479527115822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081389039754868 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081389039754868, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081389039754868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12038353830575943, + 0.03513099625706673, + 0.14034521579742432, + 0.033483389765024185, + 0.05457320064306259, + 0.02596275508403778, + 0.02493244782090187, + 0.043711233884096146, + 0.026681439951062202, + 0.4947958290576935 + ], + "credit_map": [ + 0.014108487404882908, + 0.019323429092764854, + 0.014356433413922787, + 0.012508860789239407, + 0.009079689159989357, + 0.009317627176642418, + 0.023664411157369614, + 0.018024588003754616, + 0.004733241628855467, + 0.004690422676503658, + 0.004371164366602898, + 0.004820996429771185, + 0.004862911067903042, + 0.0032732076942920685, + 0.004816106520593166, + 0.003562942147254944, + 0.012060664594173431, + 0.01553172804415226, + 0.018689097836613655, + 0.016067152842879295, + 0.03139084577560425, + 0.018214890733361244, + 0.015345416963100433, + 0.013045405969023705, + 0.002650820417329669, + 0.003102906048297882, + 0.003801078302785754, + 0.007065139710903168, + 0.003947177901864052, + 0.0029096449725329876, + 0.005067638587206602, + 0.004938981961458921, + 0.004390569869428873, + 0.006951109040528536, + 0.006767559330910444, + 0.012471825815737247, + 0.0056565687991678715, + 0.005279583390802145, + 0.005803150124847889, + 0.007252830546349287, + 0.001958562759682536, + 0.003832197282463312, + 0.0027584026101976633, + 0.0032709515653550625, + 0.005249126814305782, + 0.0017305274959653616, + 0.0051728845573961735, + 0.00199010013602674, + 0.0018409601179882884, + 0.0018502831226214767, + 0.003277658484876156, + 0.003360951319336891, + 0.003780101193115115, + 0.002759590744972229, + 0.004173672292381525, + 0.003889229381456971, + 0.0032079978846013546, + 0.003937733359634876, + 0.007263786159455776, + 0.007368218153715134, + 0.0031729615293443203, + 0.0033078682608902454, + 0.0106615349650383, + 0.004791133105754852, + 0.0018866191385313869, + 0.002210262231528759, + 0.0027128886431455612, + 0.0044462187215685844, + 0.004961247090250254, + 0.002571097109466791, + 0.004838233347982168, + 0.0030548705253750086, + 0.016253162175416946, + 0.027937596663832664, + 0.018627885729074478, + 0.17499913275241852, + 0.01395653747022152, + 0.025695662945508957, + 0.039522964507341385, + 0.17780286073684692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.016081389039754868 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797280207276344 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797280207276344, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797280207276344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10342440009117126, + 0.06276249885559082, + 0.025332361459732056, + 0.04579966515302658, + 0.041934289038181305, + 0.07073306292295456, + 0.07106728106737137, + 0.07166043668985367, + 0.074454665184021, + 0.432831346988678 + ], + "credit_map": [ + 0.011213108897209167, + 0.014730734750628471, + 0.017163820564746857, + 0.010934429243206978, + 0.006902062799781561, + 0.008220391348004341, + 0.02112959697842598, + 0.013130256906151772, + 0.009503701701760292, + 0.007808529306203127, + 0.005893038120120764, + 0.008519130758941174, + 0.009648690931499004, + 0.009170294739305973, + 0.007523922715336084, + 0.004695191979408264, + 0.002152076456695795, + 0.0025596143677830696, + 0.0033795239869505167, + 0.0041236975230276585, + 0.003914837725460529, + 0.0035937924403697252, + 0.003248675959184766, + 0.0023601441644132137, + 0.0032585335429757833, + 0.00432585459202528, + 0.004992591682821512, + 0.008474919013679028, + 0.004748799372464418, + 0.00425031827762723, + 0.007160594686865807, + 0.008588054217398167, + 0.0040609450079500675, + 0.005165603011846542, + 0.004973349627107382, + 0.007478308863937855, + 0.005506901070475578, + 0.004118999000638723, + 0.0047532650642097, + 0.005876919254660606, + 0.005575985182076693, + 0.009935010224580765, + 0.007536448072642088, + 0.008035204373300076, + 0.015355102717876434, + 0.004303800407797098, + 0.013408081606030464, + 0.006583428010344505, + 0.004641098435968161, + 0.0057424879632890224, + 0.006693555042147636, + 0.01080403383821249, + 0.011291048489511013, + 0.007804966066032648, + 0.014627223834395409, + 0.00946287252008915, + 0.006283637136220932, + 0.006819423753768206, + 0.012799953110516071, + 0.010833940468728542, + 0.005753281991928816, + 0.006114154122769833, + 0.015874188393354416, + 0.007181859575212002, + 0.004212899599224329, + 0.005607122555375099, + 0.006805716548115015, + 0.01672184467315674, + 0.012170557864010334, + 0.006634905934333801, + 0.01142482552677393, + 0.010876789689064026, + 0.015683474019169807, + 0.024381408467888832, + 0.015558868646621704, + 0.16098830103874207, + 0.011449715122580528, + 0.02074342407286167, + 0.046844761818647385, + 0.1371813714504242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.029797280207276344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, high-contrast black-and-white overhead shot, the camera held still, focused vertically on the center of a dark cast-iron pot below. Inside, a boiling, nearly pure white broth formed a violent vortex. A whole fish—head and tail intact—tumbled uncontrollably in this turmoil. Illuminated by the harsh overhead light, the fish's slippery scales reflected dazzling white highlights, contrasting starkly with its deep black skin. Its body twisted and writhed violently with the onslaught of the scalding liquid, sometimes flung to the surface, revealing its pale underbelly, and sometimes completely engulfed by the churning bubbles and opaque broth. The fish's hollow eyes were deep black spots, its mouth slightly open, as if in a silent cry. Each tumble stirred the broth, sending its tail fin slapping violently against the surface, sending up countless bright white droplets that traced brief, sharp arcs in the air before falling back onto the inky black pot wall and the bubbling white broth. The entire picture is full of primitive and wild dynamic tension, creating an intense, tense and almost violent culinary atmosphere.", + "seed": 4763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4763_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05776265431195497, + "vlm_logprob": -0.05776265431195497, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05776265431195497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017406554892659187 + ], + "opposite_logprob_tokens": [ + -5.767406463623047 + ], + "logprob_reward": -0.017406554892659187, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.017406554892659187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1750413030385971, + 0.020374562591314316, + 0.04909173399209976, + 0.08272843807935715, + 0.030692391097545624, + 0.019987840205430984, + 0.02844364196062088, + 0.01968890056014061, + 0.02720186486840248, + 0.5467493534088135 + ], + "credit_map": [ + 0.011104175820946693, + 0.008758600801229477, + 0.07133665680885315, + 0.056869033724069595, + 0.0029441658407449722, + 0.005143794231116772, + 0.009204757399857044, + 0.009680123068392277, + 0.001976471394300461, + 0.0023574200458824635, + 0.002468619728460908, + 0.003359866561368108, + 0.0027147638611495495, + 0.0019356092670932412, + 0.0030368368607014418, + 0.002524974290281534, + 0.004109513014554977, + 0.004918668884783983, + 0.005902661941945553, + 0.007225540932267904, + 0.006612770259380341, + 0.006434496026486158, + 0.00866235326975584, + 0.005225729197263718, + 0.005815296899527311, + 0.007721961010247469, + 0.011882783845067024, + 0.0165252648293972, + 0.008406514301896095, + 0.0053275772370398045, + 0.017885232344269753, + 0.009163806214928627, + 0.0018562725745141506, + 0.0036536091938614845, + 0.004919399507343769, + 0.005760505795478821, + 0.0028344171587377787, + 0.002229114528745413, + 0.005757756996899843, + 0.0036813148763030767, + 0.0013498897897079587, + 0.001554885064251721, + 0.0021886443719267845, + 0.0023737589363008738, + 0.002668941393494606, + 0.0014966832241043448, + 0.00538678839802742, + 0.002968246815726161, + 0.0018760242965072393, + 0.0019208213780075312, + 0.0069148242473602295, + 0.0032470913138240576, + 0.0029686277266591787, + 0.002685639774426818, + 0.004117292817682028, + 0.004713319707661867, + 0.0016396524151787162, + 0.002885684836655855, + 0.0021597486920654774, + 0.0027949553914368153, + 0.002500331262126565, + 0.0018457775004208088, + 0.0031633235048502684, + 0.002699426840990782, + 0.0015186119126155972, + 0.00398835027590394, + 0.0029078617226332426, + 0.0045071751810610294, + 0.002736368216574192, + 0.0016433284617960453, + 0.006260530091822147, + 0.003639640985056758, + 0.018912894651293755, + 0.026846187189221382, + 0.19381995499134064, + 0.030663611367344856, + 0.035827770829200745, + 0.03089701198041439, + 0.015760615468025208, + 0.1940213292837143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.017406554892659187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157881081104279 + ], + "opposite_logprob_tokens": [ + -1.7015788555145264 + ], + "logprob_reward": -0.20157881081104279, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157881081104279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3058306574821472, + 0.059667471796274185, + 0.043065208941698074, + 0.026318220421671867, + 0.02948988601565361, + 0.04235481843352318, + 0.013289581052958965, + 0.015675092115998268, + 0.0598142184317112, + 0.4044947624206543 + ], + "credit_map": [ + 0.019222546368837357, + 0.01309188175946474, + 0.1256486028432846, + 0.10584357380867004, + 0.0056376708671450615, + 0.007476439233869314, + 0.015256861224770546, + 0.013653098605573177, + 0.005246518179774284, + 0.008300232701003551, + 0.00582411652430892, + 0.009053410962224007, + 0.006659447215497494, + 0.004949731752276421, + 0.009089931845664978, + 0.010544088669121265, + 0.003206713357940316, + 0.00458569498732686, + 0.00512414937838912, + 0.008030573837459087, + 0.0055748033337295055, + 0.005164655391126871, + 0.006775307934731245, + 0.004603312350809574, + 0.002140372060239315, + 0.0028089648112654686, + 0.0038066827692091465, + 0.004126058891415596, + 0.003953323699533939, + 0.001985697541385889, + 0.005012936890125275, + 0.002484185853973031, + 0.0014714279677718878, + 0.003065344877541065, + 0.003030011896044016, + 0.004668679554015398, + 0.0052713798359036446, + 0.0029090328607708216, + 0.005799328442662954, + 0.0032746850047260523, + 0.0026995583903044462, + 0.0037973688449710608, + 0.004703110549598932, + 0.004539636895060539, + 0.004700477235019207, + 0.0024204726796597242, + 0.012446509674191475, + 0.007047688122838736, + 0.0012577980523929, + 0.0012980884639546275, + 0.002611979842185974, + 0.0019146056147292256, + 0.0017919940873980522, + 0.0010695618111640215, + 0.0012940994929522276, + 0.0020514559000730515, + 0.0014971832279115915, + 0.0021846937015652657, + 0.002179261762648821, + 0.0025317587424069643, + 0.0017666892381384969, + 0.0015258232597261667, + 0.0022229389287531376, + 0.0017667451174929738, + 0.0032959356904029846, + 0.0032633631490170956, + 0.007081768941134214, + 0.01014622487127781, + 0.004803309682756662, + 0.0033244225196540356, + 0.01955791562795639, + 0.008341285400092602, + 0.0139772929251194, + 0.019636468961834908, + 0.14271946251392365, + 0.023289937525987625, + 0.026711001992225647, + 0.022802013903856277, + 0.011075291782617569, + 0.14428333938121796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.20157881081104279 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634909749031067 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.02634909749031067, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634909749031067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22155453264713287, + 0.05168793722987175, + 0.08847244828939438, + 0.06251867115497589, + 0.039213020354509354, + 0.09822552651166916, + 0.024349790066480637, + 0.19229386746883392, + 0.08750676363706589, + 0.13417746126651764 + ], + "credit_map": [ + 0.017550261691212654, + 0.026765037328004837, + 0.054805055260658264, + 0.04772345721721649, + 0.013338020071387291, + 0.012358763255178928, + 0.03297429159283638, + 0.016039639711380005, + 0.004430570174008608, + 0.005748099647462368, + 0.0054062120616436005, + 0.008152438327670097, + 0.005490818992257118, + 0.005539368372410536, + 0.009474576450884342, + 0.0074458555318415165, + 0.008890674449503422, + 0.007619233801960945, + 0.011879549361765385, + 0.015564190223813057, + 0.010626540519297123, + 0.010391836985945702, + 0.011166715994477272, + 0.012333710677921772, + 0.00388511223718524, + 0.007842769846320152, + 0.009788444265723228, + 0.013497276231646538, + 0.0069343880750238895, + 0.00449949549511075, + 0.011001545004546642, + 0.005069632548838854, + 0.003027496626600623, + 0.004559130873531103, + 0.004141231998801231, + 0.006230248138308525, + 0.004953397903591394, + 0.004935594741255045, + 0.006474781781435013, + 0.004891138523817062, + 0.006385283078998327, + 0.009916332550346851, + 0.011514241807162762, + 0.009943690150976181, + 0.012119787745177746, + 0.011882445774972439, + 0.01939919963479042, + 0.017064543440937996, + 0.002396461321040988, + 0.00213145581074059, + 0.004780014045536518, + 0.0037265720311552286, + 0.0022671993356198072, + 0.002271429169923067, + 0.0031782316509634256, + 0.0035984248388558626, + 0.015620875172317028, + 0.022850392386317253, + 0.01988881453871727, + 0.03361126035451889, + 0.02631739340722561, + 0.01864899881184101, + 0.032249681651592255, + 0.023106448352336884, + 0.004471698310226202, + 0.004949365742504597, + 0.010405300185084343, + 0.014185171574354172, + 0.007515423931181431, + 0.005022833589464426, + 0.029217397794127464, + 0.011739568784832954, + 0.005664536729454994, + 0.006570375990122557, + 0.04568032547831535, + 0.008702646940946579, + 0.010151377879083157, + 0.008437179028987885, + 0.0052699437364935875, + 0.04370107501745224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.02634909749031067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19992199540138245, + 0.056790273636579514, + 0.0986904501914978, + 0.11878706514835358, + 0.0343402624130249, + 0.0625474825501442, + 0.04496343061327934, + 0.07140059024095535, + 0.033949222415685654, + 0.27860918641090393 + ], + "credit_map": [ + 0.012086519040167332, + 0.011763195507228374, + 0.0613890178501606, + 0.06904951483011246, + 0.0073156110011041164, + 0.007049204781651497, + 0.01694132387638092, + 0.01432761549949646, + 0.006112577393651009, + 0.005796370562165976, + 0.007495360914617777, + 0.009416543878614902, + 0.005258510820567608, + 0.005808393936604261, + 0.010299712419509888, + 0.006602800451219082, + 0.013086996041238308, + 0.00905960239470005, + 0.012507610023021698, + 0.017848137766122818, + 0.007041498087346554, + 0.012185545638203621, + 0.015195738524198532, + 0.011765322647988796, + 0.0063881766982376575, + 0.011190616525709629, + 0.017441079020500183, + 0.018787626177072525, + 0.014570094645023346, + 0.00708317244425416, + 0.03168341517448425, + 0.011642876081168652, + 0.0021272480953484774, + 0.004127249587327242, + 0.0037727183662354946, + 0.005532843992114067, + 0.004149317741394043, + 0.003600877011194825, + 0.007327726110816002, + 0.0037022826727479696, + 0.00352136860601604, + 0.00801953673362732, + 0.00794215314090252, + 0.006286878604441881, + 0.008413066156208515, + 0.0071708341129124165, + 0.00791745726019144, + 0.013276186771690845, + 0.00484350835904479, + 0.002605966292321682, + 0.012147734872996807, + 0.0053079454228281975, + 0.004026176407933235, + 0.0030376953072845936, + 0.004859214648604393, + 0.008135185576975346, + 0.0064674788154661655, + 0.009323862381279469, + 0.008954077027738094, + 0.010361908935010433, + 0.008572237566113472, + 0.007805824279785156, + 0.011110657826066017, + 0.008804540149867535, + 0.0021994952112436295, + 0.0020681696478277445, + 0.003981271293014288, + 0.005969868041574955, + 0.0028933058492839336, + 0.0023102075792849064, + 0.00969136692583561, + 0.004835536237806082, + 0.009388106875121593, + 0.013891332782804966, + 0.09930287301540375, + 0.015485038980841637, + 0.01738019846379757, + 0.015191476792097092, + 0.007578161545097828, + 0.10039202123880386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.00979085173457861 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687956631183624 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.033687956631183624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687956631183624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15280751883983612, + 0.04089364409446716, + 0.031121158972382545, + 0.041244640946388245, + 0.027113698422908783, + 0.0539148673415184, + 0.018479755148291588, + 0.0768962949514389, + 0.045061662793159485, + 0.5124668478965759 + ], + "credit_map": [ + 0.009432959370315075, + 0.009516261518001556, + 0.05605305731296539, + 0.04977870360016823, + 0.004231669008731842, + 0.004821729380637407, + 0.010897327214479446, + 0.008075817488133907, + 0.005373546853661537, + 0.005242684856057167, + 0.005370333790779114, + 0.007234775461256504, + 0.003907236270606518, + 0.00373462843708694, + 0.005560083780437708, + 0.004470352083444595, + 0.002186338184401393, + 0.004219342954456806, + 0.0038116287905722857, + 0.005476988852024078, + 0.0034873890690505505, + 0.003185498993843794, + 0.00497559504583478, + 0.0037783763837069273, + 0.0023158909752964973, + 0.004585204180330038, + 0.004896063357591629, + 0.006754701025784016, + 0.00512084923684597, + 0.0030465135350823402, + 0.010013665072619915, + 0.0045117540284991264, + 0.002476614201441407, + 0.002987293526530266, + 0.0036076109390705824, + 0.0037998356856405735, + 0.0026872223243117332, + 0.002777380170300603, + 0.005886398255825043, + 0.0028913430869579315, + 0.00433774571865797, + 0.008678034879267216, + 0.0072053344920277596, + 0.006536838598549366, + 0.00744976568967104, + 0.0040324777364730835, + 0.008685599081218243, + 0.006989069748669863, + 0.0019407439976930618, + 0.0017527226591482759, + 0.003954344429075718, + 0.002212163060903549, + 0.0018594468710944057, + 0.0017867316491901875, + 0.0021362837869673967, + 0.002837318228557706, + 0.004701754078269005, + 0.007731718942523003, + 0.00924927182495594, + 0.014537429437041283, + 0.009480686858296394, + 0.007705703843384981, + 0.012255777604877949, + 0.011233948171138763, + 0.002603905275464058, + 0.0029892874881625175, + 0.005262081976979971, + 0.007314816117286682, + 0.003846905892714858, + 0.0027794118504971266, + 0.01388952974230051, + 0.006375727243721485, + 0.017440324649214745, + 0.025324949994683266, + 0.18415230512619019, + 0.028557654470205307, + 0.03282215818762779, + 0.02743578515946865, + 0.012792023830115795, + 0.18394163250923157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.033687956631183624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking, high-contrast black-and-white overhead shot, the camera held still, focused vertically on the center of a dark cast-iron pot below. Inside, a boiling, nearly pure white broth formed a violent vortex. A whole fish—head and tail intact—tumbled uncontrollably in this turmoil. Illuminated by the harsh overhead light, the fish's slippery scales reflected dazzling white highlights, contrasting starkly with its deep black skin. Its body twisted and writhed violently with the onslaught of the scalding liquid, sometimes flung to the surface, revealing its pale underbelly, and sometimes completely engulfed by the churning bubbles and opaque broth. The fish's hollow eyes were deep black spots, its mouth slightly open, as if in a silent cry. Each tumble stirred the broth, sending its tail fin slapping violently against the surface, sending up countless bright white droplets that traced brief, sharp arcs in the air before falling back onto the inky black pot wall and the bubbling white broth. The entire picture is full of primitive and wild dynamic tension, creating an intense, tense and almost violent culinary atmosphere.", + "seed": 4764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_13/index_5658_seed_4764_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04035198986530304, + "vlm_logprob": -0.04035198986530304, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04035198986530304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019733885303139687 + ], + "opposite_logprob_tokens": [ + -4.769733905792236 + ], + "logprob_reward": -0.019733885303139687, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.019733885303139687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2083752304315567, + 0.07399099320173264, + 0.11477644741535187, + 0.14135196805000305, + 0.09685917943716049, + 0.04773355647921562, + 0.10408198088407516, + 0.08855584263801575, + 0.06888578832149506, + 0.05538905784487724 + ], + "credit_map": [ + 0.02504849061369896, + 0.015069961547851562, + 0.024859433993697166, + 0.0434650294482708, + 0.01605268009006977, + 0.015772687271237373, + 0.035287484526634216, + 0.032819461077451706, + 0.008415654301643372, + 0.012580732814967632, + 0.005877714138478041, + 0.009826033376157284, + 0.01022144965827465, + 0.0059607489965856075, + 0.012586170807480812, + 0.008522480726242065, + 0.0076070125214755535, + 0.009926769882440567, + 0.010758812539279461, + 0.022494690492749214, + 0.02431360073387623, + 0.007761268876492977, + 0.019057171419262886, + 0.012857111170887947, + 0.010769365355372429, + 0.01881464198231697, + 0.017519719898700714, + 0.02260914072394371, + 0.016327135264873505, + 0.010058742947876453, + 0.02423117868602276, + 0.02102203480899334, + 0.009959609247744083, + 0.0065497104078531265, + 0.006800211500376463, + 0.02476988360285759, + 0.007324391044676304, + 0.008545701391994953, + 0.008240323513746262, + 0.02466934733092785, + 0.0038010955322533846, + 0.007425243034958839, + 0.008338072337210178, + 0.009919993579387665, + 0.004143999889492989, + 0.004503341857343912, + 0.005076700355857611, + 0.004525106865912676, + 0.012760820798575878, + 0.012175077572464943, + 0.011102056130766869, + 0.012085995636880398, + 0.011803978122770786, + 0.011725095100700855, + 0.014925424009561539, + 0.01750352419912815, + 0.006848854478448629, + 0.010830296203494072, + 0.014200442470610142, + 0.013703422620892525, + 0.007963111624121666, + 0.009345506317913532, + 0.01583830639719963, + 0.009825901128351688, + 0.00577733526006341, + 0.007722705602645874, + 0.008406813256442547, + 0.009505299851298332, + 0.01208253763616085, + 0.005758788902312517, + 0.009480146691203117, + 0.010152150876820087, + 0.004235418047755957, + 0.006884019356220961, + 0.0068539888598024845, + 0.008245579898357391, + 0.00951338093727827, + 0.004082703962922096, + 0.009978477843105793, + 0.0055954838171601295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.019733885303139687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033620148897171 + ], + "opposite_logprob_tokens": [ + -2.3503363132476807 + ], + "logprob_reward": -0.10033620148897171, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033620148897171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.259459912776947, + 0.07221870869398117, + 0.07432170957326889, + 0.13703280687332153, + 0.0381021574139595, + 0.1620270013809204, + 0.06313028186559677, + 0.13297566771507263, + 0.023579547181725502, + 0.03715222328901291 + ], + "credit_map": [ + 0.027579886838793755, + 0.015831414610147476, + 0.03877440094947815, + 0.051073651760816574, + 0.02307468093931675, + 0.01746108941733837, + 0.03935648128390312, + 0.046308305114507675, + 0.006207405589520931, + 0.00944309588521719, + 0.01058278139680624, + 0.008960343897342682, + 0.010550258681178093, + 0.006191593129187822, + 0.012790541164577007, + 0.007492687087506056, + 0.006956279277801514, + 0.006180243566632271, + 0.008269027806818485, + 0.013726833276450634, + 0.009594656527042389, + 0.006592902354896069, + 0.012310139834880829, + 0.010691629722714424, + 0.008249543607234955, + 0.01452354434877634, + 0.02518012933433056, + 0.017772208899259567, + 0.01572740077972412, + 0.00843233522027731, + 0.021842969581484795, + 0.02530466951429844, + 0.004006018862128258, + 0.007358945906162262, + 0.0037826206535100937, + 0.005110055673867464, + 0.00440116785466671, + 0.0026552744675427675, + 0.006534466985613108, + 0.0042536091059446335, + 0.013154695741832256, + 0.03419547900557518, + 0.01024729572236538, + 0.023275941610336304, + 0.012790851294994354, + 0.016580279916524887, + 0.015890760347247124, + 0.03589169308543205, + 0.005556943826377392, + 0.007694303523749113, + 0.008059044368565083, + 0.011755219660699368, + 0.006392789538949728, + 0.004967742599546909, + 0.010528112761676311, + 0.008176121860742569, + 0.010964111424982548, + 0.017078621312975883, + 0.02345461957156658, + 0.0160788856446743, + 0.014866023324429989, + 0.009627277962863445, + 0.024882832542061806, + 0.01602330058813095, + 0.0026028421707451344, + 0.00245595071464777, + 0.0023382240906357765, + 0.0030158739537000656, + 0.005408705212175846, + 0.0019569136202335358, + 0.0034686168655753136, + 0.0023324231151491404, + 0.0019378227880224586, + 0.005946462508291006, + 0.0060185762122273445, + 0.006461422424763441, + 0.0034838500432670116, + 0.00205805036239326, + 0.007838526740670204, + 0.0034075125586241484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.10033620148897171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370178863406181 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.03370178863406181, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370178863406181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10583259165287018, + 0.25167420506477356, + 0.05231020599603653, + 0.145049586892128, + 0.10482338815927505, + 0.038341913372278214, + 0.04288250952959061, + 0.11903125792741776, + 0.027356546372175217, + 0.11269776523113251 + ], + "credit_map": [ + 0.009722698479890823, + 0.0060935430228710175, + 0.012636618688702583, + 0.014295299537479877, + 0.011027171276509762, + 0.00867895781993866, + 0.016996517777442932, + 0.02638179436326027, + 0.02597372606396675, + 0.03710201755166054, + 0.032811328768730164, + 0.033204980194568634, + 0.040414098650217056, + 0.017184605821967125, + 0.04172781854867935, + 0.023255661129951477, + 0.003271580208092928, + 0.003931787330657244, + 0.0050621191039681435, + 0.013575057499110699, + 0.011107751168310642, + 0.002842601854354143, + 0.009217951446771622, + 0.0033013641368597746, + 0.00875736866146326, + 0.01701022870838642, + 0.01606602780520916, + 0.021433506160974503, + 0.02450716495513916, + 0.012191726826131344, + 0.028386786580085754, + 0.016696786507964134, + 0.009821386076509953, + 0.009972510859370232, + 0.007023132871836424, + 0.020817190408706665, + 0.00956968404352665, + 0.009681019000709057, + 0.011533292010426521, + 0.026405178010463715, + 0.0027810907922685146, + 0.009484893642365932, + 0.002912836382165551, + 0.006281862501055002, + 0.002604349283501506, + 0.002947816625237465, + 0.0032196289394050837, + 0.00810944102704525, + 0.005137049127370119, + 0.007015453651547432, + 0.0043752421624958515, + 0.005328588653355837, + 0.004469350446015596, + 0.004352616611868143, + 0.007045707665383816, + 0.005158502142876387, + 0.010286914184689522, + 0.014639846980571747, + 0.025702567771077156, + 0.015246910974383354, + 0.013013051822781563, + 0.008784104138612747, + 0.01980605721473694, + 0.011551812291145325, + 0.002118278294801712, + 0.0028231986798346043, + 0.0030350876040756702, + 0.0026362044736742973, + 0.007481195032596588, + 0.0020522908307611942, + 0.004390396643429995, + 0.0028198971413075924, + 0.0061823539435863495, + 0.014164799824357033, + 0.017770417034626007, + 0.01637796312570572, + 0.02110578678548336, + 0.008348947390913963, + 0.019529957324266434, + 0.00921755749732256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.03370178863406181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181990098208189 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.0181990098208189, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181990098208189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13962818682193756, + 0.1459943801164627, + 0.06505631655454636, + 0.08329697698354721, + 0.0870557352900505, + 0.0843837782740593, + 0.190981924533844, + 0.11749401688575745, + 0.014788209460675716, + 0.0713205337524414 + ], + "credit_map": [ + 0.009142193011939526, + 0.011419053189456463, + 0.01938503235578537, + 0.029048098251223564, + 0.014524353668093681, + 0.01013690885156393, + 0.025612009689211845, + 0.020360538735985756, + 0.011045022867619991, + 0.026780061423778534, + 0.016380947083234787, + 0.020840877667069435, + 0.025724582374095917, + 0.008013886399567127, + 0.02587859332561493, + 0.011330410838127136, + 0.003479856066405773, + 0.004809254314750433, + 0.0060106306336820126, + 0.01869083382189274, + 0.009333164431154728, + 0.0035153906792402267, + 0.014116422273218632, + 0.0051007624715566635, + 0.00509648909792304, + 0.010152029804885387, + 0.008699262514710426, + 0.013793833553791046, + 0.01416796911507845, + 0.007148704957216978, + 0.016323087736964226, + 0.007915593683719635, + 0.008830942213535309, + 0.00955738965421915, + 0.007490681018680334, + 0.016411984339356422, + 0.009627513587474823, + 0.0077745565213263035, + 0.010595859959721565, + 0.016766807064414024, + 0.007097545079886913, + 0.014666921459138393, + 0.006717256736010313, + 0.010540802963078022, + 0.006436951458454132, + 0.009079388342797756, + 0.008477596566081047, + 0.02136731706559658, + 0.020633025094866753, + 0.023795880377292633, + 0.02120748534798622, + 0.030203543603420258, + 0.023661095649003983, + 0.01696361042559147, + 0.02928142435848713, + 0.0252358578145504, + 0.010388243943452835, + 0.015250272117555141, + 0.023957470431923866, + 0.015483137220144272, + 0.01291507389396429, + 0.008660392835736275, + 0.018382981419563293, + 0.012456447817385197, + 0.0017618624260649085, + 0.0014922987902536988, + 0.0019843149930238724, + 0.0015501711750403047, + 0.0028625044506043196, + 0.001244171755388379, + 0.0022979192435741425, + 0.0015949672088027, + 0.0050456528551876545, + 0.00939838495105505, + 0.011134953238070011, + 0.0092127975076437, + 0.012274912558495998, + 0.005913656670600176, + 0.011148173362016678, + 0.007192005403339863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0181990098208189 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-contrast black-and-white overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still and focused vertically on a dark cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a whole fish with head and tail intact visible tumbling in a boiling white broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fish's scales reflect dazzling white highlights against its deep black skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright white droplets seen flying from the fish's tail fin against the inky black pot wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789064079523087 + ], + "opposite_logprob_tokens": [ + -3.5297889709472656 + ], + "logprob_reward": -0.029789064079523087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789064079523087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11424411833286285, + 0.12763087451457977, + 0.026467639952898026, + 0.05377000570297241, + 0.17600519955158234, + 0.09131418913602829, + 0.18901750445365906, + 0.091305673122406, + 0.02656714804470539, + 0.1036776453256607 + ], + "credit_map": [ + 0.01663915254175663, + 0.008884995244443417, + 0.01623227447271347, + 0.021467233076691628, + 0.010152784176170826, + 0.008788606151938438, + 0.01790025271475315, + 0.014178816229104996, + 0.007752686273306608, + 0.025446970015764236, + 0.015468019992113113, + 0.01853255182504654, + 0.017746509984135628, + 0.007285906933248043, + 0.025996731594204903, + 0.00940149649977684, + 0.0035035908222198486, + 0.00267835333943367, + 0.0024629973340779543, + 0.00579844880849123, + 0.0030765181872993708, + 0.0027475506067276, + 0.00397255877032876, + 0.0022276209201663733, + 0.004103533457964659, + 0.006556820124387741, + 0.004844082053750753, + 0.007260398473590612, + 0.009664805606007576, + 0.0055716754868626595, + 0.008615085855126381, + 0.007153601851314306, + 0.015378881245851517, + 0.015812229365110397, + 0.012400809675455093, + 0.038223009556531906, + 0.01640021800994873, + 0.015557719394564629, + 0.016667423769831657, + 0.045564915984869, + 0.008326866663992405, + 0.020739929750561714, + 0.006931919604539871, + 0.011979403905570507, + 0.006301160901784897, + 0.010062113404273987, + 0.008047196082770824, + 0.01892559975385666, + 0.01937183551490307, + 0.017545035108923912, + 0.026459818705916405, + 0.028944559395313263, + 0.01840568520128727, + 0.016531582921743393, + 0.033544935286045074, + 0.028214050456881523, + 0.008552584797143936, + 0.013852044939994812, + 0.013796867802739143, + 0.011687787249684334, + 0.009219394996762276, + 0.008626355789601803, + 0.015207807533442974, + 0.010362830944359303, + 0.00279664876870811, + 0.002771761268377304, + 0.003055507317185402, + 0.0029261745512485504, + 0.00571731012314558, + 0.0018072081729769707, + 0.004303047899156809, + 0.003189489943906665, + 0.006290033459663391, + 0.014566964469850063, + 0.01786677911877632, + 0.01587522402405739, + 0.013811052776873112, + 0.006516364403069019, + 0.019176432862877846, + 0.009574797935783863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.029789064079523087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_tensor([1380, 1380])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_tensor([1380, 1380])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..05d3d8191c5edb5a1664f724bdc0f08322fccd40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_tensor([1380, 1380])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot from behind captures an elderly, balding African-American woman, the camera moving smoothly as she rides at a slow, relaxed pace. She sits poised and comfortable atop a handsome chestnut horse with a flowing black mane, her body swaying gently to the animal's steady gait. She wears classic, slightly faded blue jeans and simple, well-worn brown leather riding boots, topped by a loose-fitting white cotton shirt that flutters gently in the breeze. They travel along a wide dirt path that winds through a vibrant park. Bright, warm sunlight filters through the fresh lime-green leaves of the canopy trees, casting ever-changing patterns of light and shadow. The background reveals lush lawns and other budding foliage, creating a rich tapestry of spring colors. The entire scene is imbued with an atmosphere of tranquility and serenity, centered on the quiet, contemplative connection between the rider and her horse amidst the beauty of a perfect day.", + "seed": 13924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_1380_seed_13924_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036156623438000676, + "vlm_logprob": -0.036156623438000676, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.036156623438000676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02395193837583065 + ], + "opposite_logprob_tokens": [ + -4.648952007293701 + ], + "logprob_reward": -0.02395193837583065, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02395193837583065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055185846984386444, + 0.09147416800260544, + 0.2782166004180908, + 0.040086306631565094, + 0.16090096533298492, + 0.2141592651605606, + 0.08108288794755936, + 0.03337182104587555, + 0.02815556898713112, + 0.017366593703627586 + ], + "credit_map": [ + 0.008855604566633701, + 0.0049463058821856976, + 0.004717675503343344, + 0.004955259617418051, + 0.005353868938982487, + 0.009204423055052757, + 0.008393318392336369, + 0.008759395219385624, + 0.008349703624844551, + 0.009848196990787983, + 0.007451183162629604, + 0.009813757613301277, + 0.014363406226038933, + 0.012592320330440998, + 0.019693033769726753, + 0.009362569078803062, + 0.05473059043288231, + 0.031195713207125664, + 0.032160885632038116, + 0.01579483039677143, + 0.03014291450381279, + 0.041974153369665146, + 0.03219037875533104, + 0.04002714529633522, + 0.005318184848874807, + 0.0044877249747514725, + 0.007509029004722834, + 0.004055484663695097, + 0.0033993639517575502, + 0.0050647868774831295, + 0.00579680735245347, + 0.0044549256563186646, + 0.01828727126121521, + 0.01618676632642746, + 0.009203068912029266, + 0.023200346156954765, + 0.022067872807383537, + 0.03575518727302551, + 0.020371437072753906, + 0.015829017385840416, + 0.02576887607574463, + 0.019770456477999687, + 0.01438660267740488, + 0.028265822678804398, + 0.03846201300621033, + 0.02636757679283619, + 0.03419376164674759, + 0.02694416604936123, + 0.007779709994792938, + 0.0099021689966321, + 0.004094026982784271, + 0.010059273801743984, + 0.01862538792192936, + 0.011087643913924694, + 0.010092874057590961, + 0.00944179855287075, + 0.003923059441149235, + 0.003838985925540328, + 0.002612789859995246, + 0.00645052595064044, + 0.004952674265950918, + 0.0036163777112960815, + 0.00282572815194726, + 0.005151681136339903, + 0.003236090997233987, + 0.004946048371493816, + 0.002551350509747863, + 0.00382030475884676, + 0.004571741446852684, + 0.00395837239921093, + 0.0022695749066770077, + 0.0028020846657454967, + 0.0022995448671281338, + 0.0036465211305767298, + 0.0018837156239897013, + 0.0019062993815168738, + 0.0020966301672160625, + 0.002104488667100668, + 0.0016477477038279176, + 0.0017816465115174651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.02395193837583065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10040262341499329 + ], + "opposite_logprob_tokens": [ + -2.350402593612671 + ], + "logprob_reward": -0.10040262341499329, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10040262341499329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044983625411987305, + 0.05726558715105057, + 0.10467860102653503, + 0.0898485854268074, + 0.0671461895108223, + 0.1651207059621811, + 0.3972282409667969, + 0.05046309530735016, + 0.014196235686540604, + 0.009069071151316166 + ], + "credit_map": [ + 0.006032987963408232, + 0.005744864698499441, + 0.004966863431036472, + 0.003697781590744853, + 0.005595673341304064, + 0.006260416470468044, + 0.007548751309514046, + 0.005136291962116957, + 0.008785504847764969, + 0.005056228023022413, + 0.0049913423135876656, + 0.006254041567444801, + 0.006958161946386099, + 0.008096423000097275, + 0.010804948396980762, + 0.006318944972008467, + 0.02048371359705925, + 0.010804001241922379, + 0.010354714468121529, + 0.007030128501355648, + 0.010649188421666622, + 0.015028651803731918, + 0.01720014400780201, + 0.013128072954714298, + 0.011212063021957874, + 0.011856898665428162, + 0.014988056384027004, + 0.012357786297798157, + 0.008407327346503735, + 0.011016433127224445, + 0.01175007689744234, + 0.008259962312877178, + 0.008544228039681911, + 0.008195143193006516, + 0.0036410007160156965, + 0.009236614219844341, + 0.008914519101381302, + 0.011831956915557384, + 0.008807570673525333, + 0.007975169457495213, + 0.019630927592515945, + 0.020248129963874817, + 0.010540449991822243, + 0.021702151745557785, + 0.026713769882917404, + 0.021387767046689987, + 0.020372994244098663, + 0.024524541571736336, + 0.03922577202320099, + 0.0500805489718914, + 0.019453706219792366, + 0.056676190346479416, + 0.08350452035665512, + 0.052915602922439575, + 0.043490588665008545, + 0.051881350576877594, + 0.005371203646063805, + 0.01022859662771225, + 0.004296030383557081, + 0.009412282146513462, + 0.0062918029725551605, + 0.005766803398728371, + 0.0034763996955007315, + 0.005619985517114401, + 0.001373099978081882, + 0.001717304578050971, + 0.0020521150436252356, + 0.0018677156185731292, + 0.002490940736606717, + 0.0022510576527565718, + 0.0012192343128845096, + 0.0012247704435139894, + 0.0010367665672674775, + 0.002504018135368824, + 0.0005996086983941495, + 0.0011387544218450785, + 0.0013184010749682784, + 0.001065724529325962, + 0.0006591017008759081, + 0.0007466975366696715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.10040262341499329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254606619477272 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.01254606619477272, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254606619477272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08992847055196762, + 0.14692898094654083, + 0.23778904974460602, + 0.08921871334314346, + 0.12159189581871033, + 0.11845272034406662, + 0.08052396029233932, + 0.0659375935792923, + 0.031923808157444, + 0.017704708501696587 + ], + "credit_map": [ + 0.013145240023732185, + 0.01097512524574995, + 0.011618949472904205, + 0.009052443318068981, + 0.007893824018537998, + 0.012782675214111805, + 0.013507600873708725, + 0.01095262449234724, + 0.015938661992549896, + 0.014820856973528862, + 0.013534100726246834, + 0.01709897816181183, + 0.021422743797302246, + 0.01969190128147602, + 0.031633563339710236, + 0.012788204476237297, + 0.05372298136353493, + 0.016570914536714554, + 0.021831020712852478, + 0.014433428645133972, + 0.02040313556790352, + 0.030899973586201668, + 0.04514849931001663, + 0.03477911278605461, + 0.012460865080356598, + 0.007593047339469194, + 0.013603649102151394, + 0.010680572129786015, + 0.005992015358060598, + 0.013532563112676144, + 0.015749191865324974, + 0.009606821462512016, + 0.01487746648490429, + 0.017497941851615906, + 0.009658022783696651, + 0.013219349086284637, + 0.016690844669938087, + 0.0235725287348032, + 0.012747853994369507, + 0.01332790032029152, + 0.01921776682138443, + 0.013066074810922146, + 0.011227242648601532, + 0.013462236151099205, + 0.010838773101568222, + 0.016911935061216354, + 0.013779337517917156, + 0.019949370995163918, + 0.01167284045368433, + 0.010978694073855877, + 0.0067710280418396, + 0.01467584166675806, + 0.011748643592000008, + 0.009823691099882126, + 0.006127639673650265, + 0.008725590072572231, + 0.0077876802533864975, + 0.014949027448892593, + 0.005703743547201157, + 0.012310443446040154, + 0.007047590799629688, + 0.007584328297525644, + 0.004719825461506844, + 0.005834964103996754, + 0.0039920443668961525, + 0.0032979026436805725, + 0.004095851443707943, + 0.0036334439646452665, + 0.006274245213717222, + 0.004093441180884838, + 0.0028709738980978727, + 0.0036659117322415113, + 0.001950784120708704, + 0.0034416241105645895, + 0.0014128031907603145, + 0.0023784374352544546, + 0.0029697311110794544, + 0.0018943818286061287, + 0.0014061053516343236, + 0.0022508432157337666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.01254606619477272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301932960748672 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.023301932960748672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301932960748672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059652991592884064, + 0.06186607852578163, + 0.44062912464141846, + 0.03468024730682373, + 0.0628802701830864, + 0.11217743158340454, + 0.12540002167224884, + 0.0556427426636219, + 0.031981609761714935, + 0.015089516527950764 + ], + "credit_map": [ + 0.006175843998789787, + 0.0066048260778188705, + 0.008567011915147305, + 0.0071898456662893295, + 0.005319076590240002, + 0.009255090728402138, + 0.00833373423665762, + 0.008207560516893864, + 0.011302752420306206, + 0.004664061591029167, + 0.005436512641608715, + 0.003937665373086929, + 0.005773500073701143, + 0.01009816862642765, + 0.013250049203634262, + 0.007403368595987558, + 0.08539142459630966, + 0.04679486155509949, + 0.04815719649195671, + 0.02777949906885624, + 0.0491383858025074, + 0.06764549016952515, + 0.052967824041843414, + 0.06275442242622375, + 0.0053503550589084625, + 0.0037336417008191347, + 0.003688345430418849, + 0.0049944776110351086, + 0.0027494048699736595, + 0.004596288315951824, + 0.004728258587419987, + 0.0048394761979579926, + 0.006297209765762091, + 0.008234581910073757, + 0.006391951814293861, + 0.007504538632929325, + 0.008141144178807735, + 0.013300660997629166, + 0.007156196981668472, + 0.005853987764567137, + 0.012451074086129665, + 0.01899888552725315, + 0.009787618182599545, + 0.016896577551960945, + 0.013626707717776299, + 0.01472220104187727, + 0.011268522590398788, + 0.014425849542021751, + 0.01345799956470728, + 0.020171767100691795, + 0.00862225890159607, + 0.017902018502354622, + 0.023108920082449913, + 0.014404360204935074, + 0.011828410439193249, + 0.01590428128838539, + 0.0055657983757555485, + 0.01205777283757925, + 0.0045899380929768085, + 0.010646789334714413, + 0.006684518419206142, + 0.006295927334576845, + 0.0040923417545855045, + 0.005709656048566103, + 0.003406246192753315, + 0.006191475782543421, + 0.003315713256597519, + 0.004581353161484003, + 0.004287757910788059, + 0.004505090415477753, + 0.002483121119439602, + 0.00321085168980062, + 0.0014685175847262144, + 0.003922810312360525, + 0.0009917895076796412, + 0.001569683663547039, + 0.002190530998632312, + 0.001901127747260034, + 0.0009942309698089957, + 0.002050824696198106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.023301932960748672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580556243658066 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580556243658066, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580556243658066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05306731536984444, + 0.05594489723443985, + 0.22821569442749023, + 0.10692869126796722, + 0.19345805048942566, + 0.10709214210510254, + 0.09944646805524826, + 0.09698987752199173, + 0.024791371077299118, + 0.03406551852822304 + ], + "credit_map": [ + 0.0070219701156020164, + 0.005283991806209087, + 0.007806375157088041, + 0.0074448902159929276, + 0.003618580987676978, + 0.00787428393959999, + 0.007653114385902882, + 0.006364108994603157, + 0.007122003007680178, + 0.004483276512473822, + 0.005684785079210997, + 0.00529609527438879, + 0.006966003216803074, + 0.007889787666499615, + 0.011652356013655663, + 0.006850594189018011, + 0.04846208542585373, + 0.02536274865269661, + 0.025558684021234512, + 0.012780406512320042, + 0.02177792228758335, + 0.03506382182240486, + 0.026583034545183182, + 0.03262699022889137, + 0.016493545845150948, + 0.014184524305164814, + 0.008557410910725594, + 0.013079333119094372, + 0.0072864205576479435, + 0.014905230142176151, + 0.01849086955189705, + 0.013931361958384514, + 0.02497771568596363, + 0.027899742126464844, + 0.019359951838850975, + 0.027399517595767975, + 0.01872101239860058, + 0.030426817014813423, + 0.022071940824389458, + 0.022601358592510223, + 0.01397053711116314, + 0.007866390980780125, + 0.006234344560652971, + 0.009223460219800472, + 0.024486904963850975, + 0.015436332672834396, + 0.012654593214392662, + 0.017219576984643936, + 0.014679337851703167, + 0.008893948048353195, + 0.007121327333152294, + 0.008900045417249203, + 0.021199699491262436, + 0.014843507669866085, + 0.010532502084970474, + 0.013276094570755959, + 0.011004810221493244, + 0.021939778700470924, + 0.00717575941234827, + 0.018837712705135345, + 0.010068233124911785, + 0.011063240468502045, + 0.007255914155393839, + 0.009644434787333012, + 0.002995183924213052, + 0.002426804043352604, + 0.003124284092336893, + 0.0024531965609639883, + 0.005142535548657179, + 0.004269380122423172, + 0.0019728585612028837, + 0.0024071303196251392, + 0.0033937059342861176, + 0.008052226155996323, + 0.0033500809222459793, + 0.00415181927382946, + 0.004737220238894224, + 0.004045265261083841, + 0.0025207551661878824, + 0.003814446972683072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.020580556243658066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot from behind captures an elderly, balding African-American woman, the camera moving smoothly as she rides at a slow, relaxed pace. She sits poised and comfortable atop a handsome chestnut horse with a flowing black mane, her body swaying gently to the animal's steady gait. She wears classic, slightly faded blue jeans and simple, well-worn brown leather riding boots, topped by a loose-fitting white cotton shirt that flutters gently in the breeze. They travel along a wide dirt path that winds through a vibrant park. Bright, warm sunlight filters through the fresh lime-green leaves of the canopy trees, casting ever-changing patterns of light and shadow. The background reveals lush lawns and other budding foliage, creating a rich tapestry of spring colors. The entire scene is imbued with an atmosphere of tranquility and serenity, centered on the quiet, contemplative connection between the rider and her horse amidst the beauty of a perfect day.", + "seed": 13925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_1380_seed_13925_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06245620884001255, + "vlm_logprob": -0.06245620884001255, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06245620884001255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0567072294652462 + ], + "opposite_logprob_tokens": [ + -3.8067071437835693 + ], + "logprob_reward": -0.0567072294652462, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0567072294652462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23013626039028168, + 0.09115149080753326, + 0.15526701509952545, + 0.053368497639894485, + 0.08052211254835129, + 0.08754297345876694, + 0.07815208286046982, + 0.05013697221875191, + 0.07936808466911316, + 0.0943545550107956 + ], + "credit_map": [ + 0.0381648987531662, + 0.03754834830760956, + 0.029140302911400795, + 0.033754613250494, + 0.018604416400194168, + 0.037376612424850464, + 0.018933208659291267, + 0.016613854095339775, + 0.01518896222114563, + 0.010645187459886074, + 0.006298265885561705, + 0.013972551561892033, + 0.006195101421326399, + 0.019639093428850174, + 0.010801847092807293, + 0.008410465903580189, + 0.024565065279603004, + 0.016341904178261757, + 0.017430627718567848, + 0.011316646821796894, + 0.020569656044244766, + 0.030893536284565926, + 0.017385881394147873, + 0.016763681545853615, + 0.009300761856138706, + 0.0055626919493079185, + 0.005843083839863539, + 0.004273734055459499, + 0.006467369385063648, + 0.010412756353616714, + 0.006255057640373707, + 0.0052530341781675816, + 0.016257597133517265, + 0.007381298579275608, + 0.013326852582395077, + 0.0075860703364014626, + 0.007705948315560818, + 0.012393123470246792, + 0.007037624251097441, + 0.008833589032292366, + 0.009846968576312065, + 0.008001498878002167, + 0.009244468063116074, + 0.0071188160218298435, + 0.007589873392134905, + 0.013117283582687378, + 0.01335300412029028, + 0.01927105151116848, + 0.012563707306981087, + 0.007146233227103949, + 0.0097101591527462, + 0.0057464963756501675, + 0.0074670878238976, + 0.01033796276897192, + 0.01660098508000374, + 0.008579441346228123, + 0.006474674679338932, + 0.004917220212519169, + 0.00589285884052515, + 0.004709654487669468, + 0.006763600744307041, + 0.005968452896922827, + 0.0053277271799743176, + 0.01008277852088213, + 0.010877378284931183, + 0.009648120030760765, + 0.008916724473237991, + 0.0053598470985889435, + 0.0070546893402934074, + 0.011288603767752647, + 0.012944687157869339, + 0.013278019614517689, + 0.021639667451381683, + 0.011029892601072788, + 0.008525864221155643, + 0.0047812312841415405, + 0.006816759239882231, + 0.017464198172092438, + 0.013247586786746979, + 0.010849349200725555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0567072294652462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16045330464839935 + ], + "opposite_logprob_tokens": [ + -1.9104533195495605 + ], + "logprob_reward": -0.16045330464839935, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16045330464839935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14101384580135345, + 0.06685399264097214, + 0.10405948758125305, + 0.06731528043746948, + 0.09812642633914948, + 0.09231512248516083, + 0.15467020869255066, + 0.14431414008140564, + 0.04086028411984444, + 0.09047120809555054 + ], + "credit_map": [ + 0.015937859192490578, + 0.021253205835819244, + 0.013394218869507313, + 0.020391853526234627, + 0.012100959196686745, + 0.017822349444031715, + 0.017568569630384445, + 0.02254483662545681, + 0.008509698323905468, + 0.009506468661129475, + 0.008002480491995811, + 0.007455732673406601, + 0.006468059029430151, + 0.010499485768377781, + 0.008919350802898407, + 0.007492723409086466, + 0.010689127258956432, + 0.009627830237150192, + 0.009514572098851204, + 0.008189236745238304, + 0.022357773035764694, + 0.01786666177213192, + 0.01447157934308052, + 0.011342708952724934, + 0.008726745843887329, + 0.007385376840829849, + 0.006383409257978201, + 0.007186130154877901, + 0.007724755443632603, + 0.011738835833966732, + 0.009306749328970909, + 0.008863282389938831, + 0.012743553146719933, + 0.012356908060610294, + 0.009921055287122726, + 0.00983291957527399, + 0.015299403108656406, + 0.013152185827493668, + 0.013668330386281013, + 0.011152075603604317, + 0.015972275286912918, + 0.0105177266523242, + 0.014640498906373978, + 0.006913274060934782, + 0.008207341656088829, + 0.014041767455637455, + 0.010388177819550037, + 0.01163406204432249, + 0.01820690743625164, + 0.017617153003811836, + 0.012881200760602951, + 0.014130661264061928, + 0.021041739732027054, + 0.018842365592718124, + 0.03501526266336441, + 0.016934938728809357, + 0.020337959751486778, + 0.01756591722369194, + 0.010099704377353191, + 0.014771843329071999, + 0.0208108089864254, + 0.019361190497875214, + 0.01612105965614319, + 0.02524566277861595, + 0.008061784319579601, + 0.00509248161688447, + 0.004499255213886499, + 0.0023839978966861963, + 0.003006270620971918, + 0.0059315478429198265, + 0.0033782098907977343, + 0.008506739512085915, + 0.015385138802230358, + 0.009065661579370499, + 0.009698730893433094, + 0.004275842569768429, + 0.00857302825897932, + 0.015839694067835808, + 0.012885300442576408, + 0.014747818000614643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.16045330464839935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304378613829613 + ], + "opposite_logprob_tokens": [ + -3.7733044624328613 + ], + "logprob_reward": -0.023304378613829613, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304378613829613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1963314414024353, + 0.09720417112112045, + 0.15420132875442505, + 0.05937451496720314, + 0.07672496140003204, + 0.17049479484558105, + 0.10599269717931747, + 0.04618171602487564, + 0.03381190448999405, + 0.05968251824378967 + ], + "credit_map": [ + 0.030982764437794685, + 0.027743065729737282, + 0.019468633458018303, + 0.021364063024520874, + 0.01815573312342167, + 0.0393485389649868, + 0.01636328175663948, + 0.02290535345673561, + 0.012661119922995567, + 0.012006782926619053, + 0.008688663132488728, + 0.014729388989508152, + 0.007317657582461834, + 0.016058845445513725, + 0.01271267980337143, + 0.013029035180807114, + 0.018334805965423584, + 0.019107114523649216, + 0.014812818728387356, + 0.012312310747802258, + 0.029994206503033638, + 0.023850666359066963, + 0.013326331973075867, + 0.02246306650340557, + 0.010184071958065033, + 0.0070235575549304485, + 0.00869093369692564, + 0.004455810412764549, + 0.007563972845673561, + 0.008730416186153889, + 0.006201580632477999, + 0.006524169817566872, + 0.010094482451677322, + 0.009233275428414345, + 0.009087502025067806, + 0.006598545238375664, + 0.012053067795932293, + 0.011309501715004444, + 0.008312528021633625, + 0.010036060586571693, + 0.0174696184694767, + 0.019106198102235794, + 0.01290749292820692, + 0.01310151070356369, + 0.01896505244076252, + 0.023857761174440384, + 0.03084280714392662, + 0.0342443585395813, + 0.017233068123459816, + 0.012446984648704529, + 0.01307451818138361, + 0.008695954456925392, + 0.010867281816899776, + 0.013532225042581558, + 0.018456924706697464, + 0.011685751378536224, + 0.007413729093968868, + 0.005990231409668922, + 0.0064816889353096485, + 0.004379084333777428, + 0.005688054487109184, + 0.005880424752831459, + 0.0063465735875070095, + 0.004001932218670845, + 0.004507324658334255, + 0.004033263772726059, + 0.003135355655103922, + 0.0017601511208340526, + 0.0028789485804736614, + 0.004392936360090971, + 0.005381574388593435, + 0.007722348440438509, + 0.014826263301074505, + 0.0066569093614816666, + 0.006488206796348095, + 0.003133433870971203, + 0.003639394883066416, + 0.010148122906684875, + 0.008694629184901714, + 0.006095559801906347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.023304378613829613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371492773294449 + ], + "opposite_logprob_tokens": [ + -3.408715009689331 + ], + "logprob_reward": -0.03371492773294449, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371492773294449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650827467441559, + 0.13375379145145416, + 0.1457280069589615, + 0.0515977144241333, + 0.07361961901187897, + 0.05289129540324211, + 0.11115962266921997, + 0.10021071881055832, + 0.01998944766819477, + 0.0459669828414917 + ], + "credit_map": [ + 0.04259810969233513, + 0.04312872886657715, + 0.03903312236070633, + 0.032267626374959946, + 0.028525106608867645, + 0.03815402090549469, + 0.021333154290914536, + 0.020042898133397102, + 0.01741015538573265, + 0.01879405789077282, + 0.012558001093566418, + 0.02925250120460987, + 0.0100167915225029, + 0.018979649990797043, + 0.014445491135120392, + 0.01229714136570692, + 0.016915157437324524, + 0.0180585328489542, + 0.013509650714695454, + 0.013212518766522408, + 0.02065489999949932, + 0.024046335369348526, + 0.021299859508872032, + 0.0180310420691967, + 0.007267792709171772, + 0.007437792606651783, + 0.0052881366573274136, + 0.004816950764507055, + 0.006421764846891165, + 0.008618338964879513, + 0.005559762939810753, + 0.006187177263200283, + 0.010206104256212711, + 0.00679327268153429, + 0.00803835503757, + 0.005439201835542917, + 0.009895914234220982, + 0.012263743206858635, + 0.010128461755812168, + 0.010854560881853104, + 0.00968784186989069, + 0.006433252710849047, + 0.006350780837237835, + 0.00571260042488575, + 0.004635500721633434, + 0.007552755530923605, + 0.004937037359923124, + 0.007581522688269615, + 0.015390793792903423, + 0.0144264567643404, + 0.008687734603881836, + 0.00900965929031372, + 0.0108522679656744, + 0.014311743900179863, + 0.02444952353835106, + 0.014031445607542992, + 0.013528644107282162, + 0.012178242206573486, + 0.00879876222461462, + 0.011362777091562748, + 0.015651371330022812, + 0.010443557985126972, + 0.014273984357714653, + 0.013973387889564037, + 0.0040435101836919785, + 0.0022896789014339447, + 0.003052757354453206, + 0.0012564942007884383, + 0.001925351214595139, + 0.002794044092297554, + 0.002122183097526431, + 0.002505428157746792, + 0.010893850587308407, + 0.004568260163068771, + 0.004548185970634222, + 0.0024399105459451675, + 0.0040824878960847855, + 0.008348766714334488, + 0.006458738353103399, + 0.004626783542335033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03371492773294449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381012037396431 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.0381012037396431, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381012037396431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08643649518489838, + 0.0502045601606369, + 0.1274077594280243, + 0.034649189561605453, + 0.10049404203891754, + 0.07871289551258087, + 0.17490418255329132, + 0.07075304538011551, + 0.09474584460258484, + 0.18169204890727997 + ], + "credit_map": [ + 0.014586701057851315, + 0.007886921055614948, + 0.012904295697808266, + 0.007982401177287102, + 0.008653405122458935, + 0.01861369051039219, + 0.007259533274918795, + 0.00854954868555069, + 0.005326492246240377, + 0.007277380209416151, + 0.005507550667971373, + 0.008490956388413906, + 0.0036644565407186747, + 0.008330605924129486, + 0.00638350797817111, + 0.005223606247454882, + 0.01955670304596424, + 0.013196298852562904, + 0.012290933169424534, + 0.00863389577716589, + 0.017732767388224602, + 0.024273967370390892, + 0.01622892916202545, + 0.015494254417717457, + 0.004104800522327423, + 0.0034006303176283836, + 0.003479941748082638, + 0.0034409372601658106, + 0.004075718577951193, + 0.008160198107361794, + 0.0039993817918002605, + 0.003987579606473446, + 0.02099095843732357, + 0.007665544748306274, + 0.019847607240080833, + 0.008882105350494385, + 0.006008240394294262, + 0.018817611038684845, + 0.006126136984676123, + 0.01215583086013794, + 0.013790108263492584, + 0.012152557261288166, + 0.014944948256015778, + 0.006071552634239197, + 0.005841732025146484, + 0.011510848067700863, + 0.006553613115102053, + 0.007847532629966736, + 0.03475066274404526, + 0.021722909063100815, + 0.035899817943573, + 0.012234979309141636, + 0.012029067613184452, + 0.02497873827815056, + 0.016890671104192734, + 0.016397330909967422, + 0.01363914180546999, + 0.0061784484423696995, + 0.014128013513982296, + 0.005335487425327301, + 0.00584205100312829, + 0.009494365192949772, + 0.006630072835832834, + 0.009505465626716614, + 0.015804773196578026, + 0.009145301766693592, + 0.01313844881951809, + 0.0053809951059520245, + 0.008458355441689491, + 0.012878451496362686, + 0.013533128425478935, + 0.016406381502747536, + 0.056771647185087204, + 0.02572738751769066, + 0.016497112810611725, + 0.008793939836323261, + 0.008845099247992039, + 0.026487456634640694, + 0.019369207322597504, + 0.01920018345117569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.0381012037396431 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot from behind captures an elderly, balding African-American woman, the camera moving smoothly as she rides at a slow, relaxed pace. She sits poised and comfortable atop a handsome chestnut horse with a flowing black mane, her body swaying gently to the animal's steady gait. She wears classic, slightly faded blue jeans and simple, well-worn brown leather riding boots, topped by a loose-fitting white cotton shirt that flutters gently in the breeze. They travel along a wide dirt path that winds through a vibrant park. Bright, warm sunlight filters through the fresh lime-green leaves of the canopy trees, casting ever-changing patterns of light and shadow. The background reveals lush lawns and other budding foliage, creating a rich tapestry of spring colors. The entire scene is imbued with an atmosphere of tranquility and serenity, centered on the quiet, contemplative connection between the rider and her horse amidst the beauty of a perfect day.", + "seed": 13926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_1380_seed_13926_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04304140321910381, + "vlm_logprob": -0.04304140321910381, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04304140321910381, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03240428864955902 + ], + "opposite_logprob_tokens": [ + -4.28240442276001 + ], + "logprob_reward": -0.03240428864955902, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03240428864955902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831457495689392, + 0.06292908638715744, + 0.03849074989557266, + 0.052571605890989304, + 0.19453831017017365, + 0.18852373957633972, + 0.06725563853979111, + 0.04841538891196251, + 0.14357542991638184, + 0.06538549065589905 + ], + "credit_map": [ + 0.013506939634680748, + 0.022851815447211266, + 0.021629314869642258, + 0.013092954643070698, + 0.012257291935384274, + 0.030197344720363617, + 0.014916451647877693, + 0.009862462989985943, + 0.005911585409194231, + 0.008696435950696468, + 0.006702806800603867, + 0.00515417056158185, + 0.009145019575953484, + 0.014690922573208809, + 0.006573664955794811, + 0.006054483354091644, + 0.00749160023406148, + 0.004172674845904112, + 0.004217357840389013, + 0.004564542789012194, + 0.006436489522457123, + 0.004134334158152342, + 0.004168327432125807, + 0.0033054205123335123, + 0.0073833167552948, + 0.008623371832072735, + 0.006263148505240679, + 0.0050253355875611305, + 0.006401400547474623, + 0.007686023134738207, + 0.005348775070160627, + 0.005840236786752939, + 0.034838177263736725, + 0.026034682989120483, + 0.038176290690898895, + 0.012711345218122005, + 0.01239820197224617, + 0.042730577290058136, + 0.011208813637495041, + 0.016440218314528465, + 0.06838749349117279, + 0.06645036488771439, + 0.009339263662695885, + 0.006159328389912844, + 0.013699477538466454, + 0.013135399669408798, + 0.004569970071315765, + 0.006782439071685076, + 0.011773192323744297, + 0.0069376942701637745, + 0.006472843699157238, + 0.004748406354337931, + 0.010279621928930283, + 0.01461180578917265, + 0.0062189362943172455, + 0.006213139742612839, + 0.007338550873100758, + 0.004519651178270578, + 0.005602838937193155, + 0.00828319787979126, + 0.005972280632704496, + 0.005871188826858997, + 0.005385289434343576, + 0.00544239254668355, + 0.023339929059147835, + 0.01610049232840538, + 0.016930853947997093, + 0.01596427708864212, + 0.029385879635810852, + 0.020880328491330147, + 0.011926572769880295, + 0.009047089144587517, + 0.009131315164268017, + 0.007757802028208971, + 0.008058030158281326, + 0.009357183240354061, + 0.011487066745758057, + 0.007591729052364826, + 0.006589202675968409, + 0.005413156468421221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.03240428864955902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908472210168839 + ], + "opposite_logprob_tokens": [ + -2.4640846252441406 + ], + "logprob_reward": -0.08908472210168839, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908472210168839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07274222373962402, + 0.033079441636800766, + 0.02026672475039959, + 0.05969933047890663, + 0.4966932237148285, + 0.13446156680583954, + 0.07568332552909851, + 0.01759464479982853, + 0.055313244462013245, + 0.034466296434402466 + ], + "credit_map": [ + 0.011090769432485104, + 0.010556220076978207, + 0.013089376501739025, + 0.006225370801985264, + 0.005613464396446943, + 0.012086469680070877, + 0.006752097513526678, + 0.007328452076762915, + 0.004101280122995377, + 0.004433637950569391, + 0.005359211936593056, + 0.003109436947852373, + 0.005119381472468376, + 0.003760833526030183, + 0.004475862719118595, + 0.0027197953313589096, + 0.0032646318431943655, + 0.002260735025629401, + 0.0029224997851997614, + 0.0022036542650312185, + 0.0023337844759225845, + 0.0026011490263044834, + 0.0022537619806826115, + 0.002426507882773876, + 0.007602876517921686, + 0.009673580527305603, + 0.005424226634204388, + 0.006245561875402927, + 0.00870133750140667, + 0.008374743163585663, + 0.0058251312002539635, + 0.007851872593164444, + 0.19488897919654846, + 0.040288373827934265, + 0.18769033253192902, + 0.01804489642381668, + 0.015702953562140465, + 0.018299415707588196, + 0.009698974899947643, + 0.012079319916665554, + 0.05042340233922005, + 0.05210449919104576, + 0.00465495465323329, + 0.0032654129900038242, + 0.010627582669258118, + 0.008073313161730766, + 0.002714181551709771, + 0.002598216524347663, + 0.010118908248841763, + 0.008301693014800549, + 0.0072180964052677155, + 0.008241239003837109, + 0.008285025134682655, + 0.008654973469674587, + 0.007859251461923122, + 0.017004143446683884, + 0.002069690730422735, + 0.0021482014562934637, + 0.0023682403843849897, + 0.0019686792511492968, + 0.0031677940860390663, + 0.0027040482964366674, + 0.0014663804322481155, + 0.0017016096971929073, + 0.007077485788613558, + 0.006658332422375679, + 0.006083587650209665, + 0.005155471619218588, + 0.011586836539208889, + 0.00973965972661972, + 0.004022903274744749, + 0.004988967906683683, + 0.006829345133155584, + 0.005079212598502636, + 0.004868614953011274, + 0.002573256380856037, + 0.003326726146042347, + 0.006355892866849899, + 0.002401283709332347, + 0.0030319637153297663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.08908472210168839 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083230376243591 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083230376243591, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083230376243591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10429243743419647, + 0.03007715381681919, + 0.0205689650028944, + 0.25501176714897156, + 0.18560312688350677, + 0.13338327407836914, + 0.0837121456861496, + 0.09098305553197861, + 0.08030495792627335, + 0.016063138842582703 + ], + "credit_map": [ + 0.011941087432205677, + 0.015356273390352726, + 0.013976712711155415, + 0.010760003700852394, + 0.010835519060492516, + 0.022065050899982452, + 0.010345780290663242, + 0.009012009017169476, + 0.003618568181991577, + 0.004557707812637091, + 0.0048018209636211395, + 0.0024506840854883194, + 0.00468325475230813, + 0.004973393864929676, + 0.0024987279903143644, + 0.002492995234206319, + 0.0034379111602902412, + 0.002627382054924965, + 0.0039059545379132032, + 0.0019043987849727273, + 0.002176430309191346, + 0.0019884733483195305, + 0.001723948516882956, + 0.0028044667560607195, + 0.03336997330188751, + 0.03875025734305382, + 0.0209835022687912, + 0.02725953236222267, + 0.04837581887841225, + 0.031120825558900833, + 0.03129567205905914, + 0.02385620027780533, + 0.04406355693936348, + 0.022875992581248283, + 0.04992639645934105, + 0.008840017952024937, + 0.009940333664417267, + 0.02946348488330841, + 0.00779307447373867, + 0.012700265273451805, + 0.0503341369330883, + 0.0509587861597538, + 0.0038964306004345417, + 0.003917061258107424, + 0.010458525270223618, + 0.007741574198007584, + 0.0027225534431636333, + 0.0033542008604854345, + 0.013361402787268162, + 0.007107229437679052, + 0.006308331154286861, + 0.008289776742458344, + 0.00865455623716116, + 0.022349439561367035, + 0.007178963627666235, + 0.010462451726198196, + 0.012114367447793484, + 0.009080651216208935, + 0.01706547662615776, + 0.011728198267519474, + 0.011782091110944748, + 0.012218363583087921, + 0.00845781248062849, + 0.008536095730960369, + 0.013545399531722069, + 0.012728814035654068, + 0.009347110986709595, + 0.011310259811580181, + 0.010005833581089973, + 0.009613940492272377, + 0.005600424483418465, + 0.008153175935149193, + 0.0016549493884667754, + 0.0020804626401513815, + 0.002168236067518592, + 0.0018270304426550865, + 0.0022125367540866137, + 0.002226250944659114, + 0.002244373317807913, + 0.0016492994036525488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.011083230376243591 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04032204672694206, + 0.02910435199737549, + 0.025161132216453552, + 0.08970848470926285, + 0.6060516238212585, + 0.053962890058755875, + 0.03650709241628647, + 0.04202009364962578, + 0.040403496474027634, + 0.03675879165530205 + ], + "credit_map": [ + 0.003983379807323217, + 0.005749677773565054, + 0.006062604952603579, + 0.00410973047837615, + 0.0035920049995183945, + 0.008743449114263058, + 0.004667440429329872, + 0.0034137561451643705, + 0.0029693064279854298, + 0.004101286642253399, + 0.003232409246265888, + 0.0019052113639190793, + 0.004578809719532728, + 0.007334481459110975, + 0.002994777634739876, + 0.0019880696199834347, + 0.005254185758531094, + 0.003158063627779484, + 0.0030774318147450686, + 0.0022793279495090246, + 0.0032037408091127872, + 0.003328645834699273, + 0.002225963631644845, + 0.0026337720919400454, + 0.015455126762390137, + 0.013938412070274353, + 0.007522095926105976, + 0.0062821777537465096, + 0.008916972205042839, + 0.0146783497184515, + 0.010133880190551281, + 0.012781478464603424, + 0.2366817742586136, + 0.04889649897813797, + 0.22656750679016113, + 0.02039545401930809, + 0.022010860964655876, + 0.024513088166713715, + 0.01191361341625452, + 0.015072833746671677, + 0.020005639642477036, + 0.020392080768942833, + 0.002661672420799732, + 0.001545981504023075, + 0.004068400245159864, + 0.0028924995567649603, + 0.0012444246094673872, + 0.0011521902633830905, + 0.005813779775053263, + 0.00345681793987751, + 0.006260788533836603, + 0.003302342724055052, + 0.004052275326102972, + 0.006353288423269987, + 0.0032088549342006445, + 0.004058944992721081, + 0.0048508718609809875, + 0.004790380597114563, + 0.004673792514950037, + 0.0056360880844295025, + 0.008622074499726295, + 0.005165648180991411, + 0.004429758060723543, + 0.0038514789193868637, + 0.005010126158595085, + 0.0064895800314843655, + 0.002944762585684657, + 0.006837363354861736, + 0.0063139344565570354, + 0.005074088927358389, + 0.0037737437523901463, + 0.003959896974265575, + 0.005115484818816185, + 0.004925537388771772, + 0.004431474022567272, + 0.004333801101893187, + 0.0033750918228179216, + 0.007275707088410854, + 0.004102046601474285, + 0.0031996453180909157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.020591532811522484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062043242156505585 + ], + "opposite_logprob_tokens": [ + -2.8120431900024414 + ], + "logprob_reward": -0.062043242156505585, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062043242156505585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06340474635362625, + 0.04064880311489105, + 0.08561788499355316, + 0.13683995604515076, + 0.23516885936260223, + 0.15376076102256775, + 0.09429731965065002, + 0.044142771512269974, + 0.12379185855388641, + 0.02232692949473858 + ], + "credit_map": [ + 0.009916897863149643, + 0.010411339811980724, + 0.011985168792307377, + 0.0051209996454417706, + 0.004994977731257677, + 0.010593115352094173, + 0.0053913043811917305, + 0.004990959540009499, + 0.004071301780641079, + 0.006292602047324181, + 0.006926564499735832, + 0.0022334582172334194, + 0.00569387711584568, + 0.01017760206013918, + 0.002589415293186903, + 0.00266398792155087, + 0.026281515136361122, + 0.013105537742376328, + 0.013186769559979439, + 0.00669054826721549, + 0.0059695858508348465, + 0.010902039706707, + 0.004305197391659021, + 0.005176702048629522, + 0.0319785512983799, + 0.023764796555042267, + 0.011482604779303074, + 0.010362723842263222, + 0.01712409034371376, + 0.020215347409248352, + 0.012169117107987404, + 0.009742740541696548, + 0.06381828337907791, + 0.032756220549345016, + 0.05027279630303383, + 0.012056976556777954, + 0.016681501641869545, + 0.03669360280036926, + 0.009581402875483036, + 0.01330812182277441, + 0.049836691468954086, + 0.04906798526644707, + 0.008870508521795273, + 0.006109778303653002, + 0.01160710584372282, + 0.01564142294228077, + 0.006338815670460463, + 0.006288473028689623, + 0.013894910924136639, + 0.00847613625228405, + 0.012525197118520737, + 0.012580539099872112, + 0.011273825541138649, + 0.016421513631939888, + 0.008074219338595867, + 0.011050987988710403, + 0.005395015235990286, + 0.005402394570410252, + 0.004590544383972883, + 0.004973097238689661, + 0.007598996628075838, + 0.005924995057284832, + 0.005052246619015932, + 0.005205487832427025, + 0.020811012014746666, + 0.014428505674004555, + 0.015869710594415665, + 0.017279190942645073, + 0.020699109882116318, + 0.013505274429917336, + 0.013469479978084564, + 0.007729580160230398, + 0.0037554209120571613, + 0.003088678466156125, + 0.004313655197620392, + 0.002777443965896964, + 0.0024279491044580936, + 0.0028086716774851084, + 0.001428929972462356, + 0.0017261812463402748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.062043242156505585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot from behind captures an elderly, balding African-American woman, the camera moving smoothly as she rides at a slow, relaxed pace. She sits poised and comfortable atop a handsome chestnut horse with a flowing black mane, her body swaying gently to the animal's steady gait. She wears classic, slightly faded blue jeans and simple, well-worn brown leather riding boots, topped by a loose-fitting white cotton shirt that flutters gently in the breeze. They travel along a wide dirt path that winds through a vibrant park. Bright, warm sunlight filters through the fresh lime-green leaves of the canopy trees, casting ever-changing patterns of light and shadow. The background reveals lush lawns and other budding foliage, creating a rich tapestry of spring colors. The entire scene is imbued with an atmosphere of tranquility and serenity, centered on the quiet, contemplative connection between the rider and her horse amidst the beauty of a perfect day.", + "seed": 13927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_14/index_1380_seed_13927_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05642062947154045, + "vlm_logprob": -0.05642062947154045, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05642062947154045, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034177809953689575 + ], + "opposite_logprob_tokens": [ + -4.159177780151367 + ], + "logprob_reward": -0.034177809953689575, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.034177809953689575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11601079255342484, + 0.05759778991341591, + 0.044140975922346115, + 0.05358051136136055, + 0.355334997177124, + 0.07534711807966232, + 0.1271020621061325, + 0.09806625545024872, + 0.047299809753894806, + 0.025519661605358124 + ], + "credit_map": [ + 0.014897500164806843, + 0.014883074909448624, + 0.024211980402469635, + 0.01187062356621027, + 0.00962498877197504, + 0.01736794225871563, + 0.011515896767377853, + 0.011638794094324112, + 0.006918476428836584, + 0.008601090870797634, + 0.006716056726872921, + 0.007784163113683462, + 0.0036414170172065496, + 0.00485894363373518, + 0.011554965749382973, + 0.00752267986536026, + 0.007328695151954889, + 0.005458884872496128, + 0.00888423528522253, + 0.004130760207772255, + 0.004074114840477705, + 0.0052170176059007645, + 0.004007714334875345, + 0.005039558745920658, + 0.005459715612232685, + 0.008393073454499245, + 0.006755408365279436, + 0.007468617055565119, + 0.006950901355594397, + 0.008418110199272633, + 0.005115380976349115, + 0.005019310861825943, + 0.10769900679588318, + 0.04931053891777992, + 0.049975208938121796, + 0.017942901700735092, + 0.016315536573529243, + 0.08777984231710434, + 0.012158653698861599, + 0.014153318479657173, + 0.013250589370727539, + 0.008022607304155827, + 0.011362412944436073, + 0.006297585088759661, + 0.007170161232352257, + 0.014102463610470295, + 0.008713223971426487, + 0.006428083870559931, + 0.01772393099963665, + 0.018734922632575035, + 0.011707636527717113, + 0.01078477781265974, + 0.023862410336732864, + 0.015294951386749744, + 0.014575829729437828, + 0.01441760640591383, + 0.015610577538609505, + 0.011963482946157455, + 0.007780481595546007, + 0.006281736772507429, + 0.019567091017961502, + 0.017371084541082382, + 0.011358283460140228, + 0.00813352596014738, + 0.0076836468651890755, + 0.00503108324483037, + 0.003830220550298691, + 0.005035750102251768, + 0.006228742655366659, + 0.00726319057866931, + 0.007266273256391287, + 0.004960907157510519, + 0.004363921470940113, + 0.00306886387988925, + 0.003105637151747942, + 0.0023099242243915796, + 0.0032618872355669737, + 0.004842014517635107, + 0.0022899052128195763, + 0.002277510007843375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.034177809953689575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299964040517807 + ], + "opposite_logprob_tokens": [ + -2.237999677658081 + ], + "logprob_reward": -0.11299964040517807, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299964040517807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13352324068546295, + 0.03881026804447174, + 0.10249896347522736, + 0.16825325787067413, + 0.16600871086120605, + 0.0610274113714695, + 0.07892261445522308, + 0.09596731513738632, + 0.04311128705739975, + 0.11187700182199478 + ], + "credit_map": [ + 0.01953163929283619, + 0.01788218319416046, + 0.023522350937128067, + 0.012074096128344536, + 0.016159994527697563, + 0.021149341017007828, + 0.012182607315480709, + 0.011021017096936703, + 0.003848383668810129, + 0.0036095266696065664, + 0.0035004769451916218, + 0.004020466469228268, + 0.007691174279898405, + 0.004303573165088892, + 0.007731571793556213, + 0.0041050985455513, + 0.024686140939593315, + 0.018267400562763214, + 0.009669688530266285, + 0.010072275064885616, + 0.009541161358356476, + 0.008561253547668457, + 0.009180084802210331, + 0.012520960532128811, + 0.022457733750343323, + 0.025501031428575516, + 0.021430809050798416, + 0.015728067606687546, + 0.027304116636514664, + 0.02042931690812111, + 0.02011609449982643, + 0.015286092646420002, + 0.04896751418709755, + 0.02094387449324131, + 0.023979617282748222, + 0.00802919827401638, + 0.009473743848502636, + 0.03996305167675018, + 0.007504356559365988, + 0.0071473619900643826, + 0.008939696475863457, + 0.008361105807125568, + 0.0074402992613613605, + 0.004885852802544832, + 0.006388251669704914, + 0.011185822077095509, + 0.00744452653452754, + 0.006381860468536615, + 0.010727102868258953, + 0.010975568555295467, + 0.00743477838113904, + 0.0073709446005523205, + 0.017308851704001427, + 0.009042227640748024, + 0.009045383892953396, + 0.007017761003226042, + 0.022075790911912918, + 0.012391342781484127, + 0.008378413505852222, + 0.005453814752399921, + 0.01520344614982605, + 0.017224140465259552, + 0.008067984133958817, + 0.0071723805740475655, + 0.008884632959961891, + 0.005501824896782637, + 0.0052305604331195354, + 0.0038437817711383104, + 0.004592095967382193, + 0.005067624617367983, + 0.005290565546602011, + 0.00470019830390811, + 0.016359195113182068, + 0.013773602433502674, + 0.010101854801177979, + 0.010755977593362331, + 0.023350952193140984, + 0.015859290957450867, + 0.011652613058686256, + 0.010023511946201324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.11299964040517807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607775315642357 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.01607775315642357, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607775315642357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15905947983264923, + 0.05143815279006958, + 0.07900059223175049, + 0.13656234741210938, + 0.1341150850057602, + 0.17181891202926636, + 0.11027146875858307, + 0.06525194644927979, + 0.06930456310510635, + 0.02317744679749012 + ], + "credit_map": [ + 0.02883267216384411, + 0.023446857929229736, + 0.03147069364786148, + 0.014019319787621498, + 0.013899088837206364, + 0.022862061858177185, + 0.012217486277222633, + 0.012311302125453949, + 0.0069768535904586315, + 0.007928255014121532, + 0.007376493886113167, + 0.006095701362937689, + 0.0060790120624005795, + 0.005856715142726898, + 0.005963847041130066, + 0.005161275155842304, + 0.012881430797278881, + 0.011338254436850548, + 0.007198815234005451, + 0.010226594284176826, + 0.011717207729816437, + 0.00680113397538662, + 0.008706233464181423, + 0.0101309260353446, + 0.014115048572421074, + 0.015075117349624634, + 0.012470537796616554, + 0.017691180109977722, + 0.027384620159864426, + 0.018783310428261757, + 0.016154926270246506, + 0.014887620694935322, + 0.025450000539422035, + 0.018662171438336372, + 0.010901886969804764, + 0.0083964504301548, + 0.021263882517814636, + 0.0244634710252285, + 0.01382037065923214, + 0.011156853288412094, + 0.02445872873067856, + 0.02156493254005909, + 0.018895473331212997, + 0.01541473064571619, + 0.02043687179684639, + 0.022098278626799583, + 0.028124384582042694, + 0.020825527608394623, + 0.01385907456278801, + 0.015807749703526497, + 0.008907404728233814, + 0.009904827922582626, + 0.024830007925629616, + 0.013053147122263908, + 0.014926307834684849, + 0.008982955478131771, + 0.014324182644486427, + 0.008385278284549713, + 0.005731834098696709, + 0.003948537167161703, + 0.010090620256960392, + 0.011268943548202515, + 0.0066780527122318745, + 0.004824504721909761, + 0.013015870936214924, + 0.006604864727705717, + 0.010844347067177296, + 0.005262990947812796, + 0.00786211434751749, + 0.00924242939800024, + 0.0091049550101161, + 0.007366994861513376, + 0.003672854509204626, + 0.0028675193898379803, + 0.0028416395653039217, + 0.0025486040394753218, + 0.0028187562711536884, + 0.0031847648788243532, + 0.002677891170606017, + 0.0025654183700680733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.01607775315642357 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813826084136963 + ], + "opposite_logprob_tokens": [ + -3.529813766479492 + ], + "logprob_reward": -0.029813826084136963, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813826084136963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17153820395469666, + 0.03291644901037216, + 0.03926548361778259, + 0.044514935463666916, + 0.05385357141494751, + 0.18096980452537537, + 0.09324529021978378, + 0.24137623608112335, + 0.030199483036994934, + 0.11212056875228882 + ], + "credit_map": [ + 0.043329738080501556, + 0.022011570632457733, + 0.02274850383400917, + 0.010926146991550922, + 0.013208656571805477, + 0.03649304807186127, + 0.010959700681269169, + 0.011860834434628487, + 0.0038949823938310146, + 0.004612696822732687, + 0.0038825897499918938, + 0.0034970969427376986, + 0.004400681238621473, + 0.00363090424798429, + 0.005283030215650797, + 0.0037144676316529512, + 0.005321035161614418, + 0.004688385408371687, + 0.004141753073781729, + 0.005349069833755493, + 0.006993045099079609, + 0.0037426624912768602, + 0.0045556072145700455, + 0.004473923705518246, + 0.0062721804715693, + 0.007788978051394224, + 0.006757964845746756, + 0.004807358141988516, + 0.004870674107223749, + 0.006022116634994745, + 0.003659085603430867, + 0.004336576908826828, + 0.013107171282172203, + 0.008537333458662033, + 0.004183698911219835, + 0.005285900551825762, + 0.00538682984188199, + 0.009952418506145477, + 0.0041198525577783585, + 0.0032803649082779884, + 0.03268963098526001, + 0.029461178928613663, + 0.02048730105161667, + 0.024510331451892853, + 0.013527401722967625, + 0.030716249719262123, + 0.018233472481369972, + 0.01134424190968275, + 0.011730884201824665, + 0.011742070317268372, + 0.007606483995914459, + 0.008764093741774559, + 0.021702077239751816, + 0.010244595818221569, + 0.012053373269736767, + 0.009401707910001278, + 0.02758175879716873, + 0.027300113812088966, + 0.015001481398940086, + 0.016765262931585312, + 0.05365034565329552, + 0.045986488461494446, + 0.032332614064216614, + 0.02275816909968853, + 0.005836402997374535, + 0.0037012274842709303, + 0.0035487902350723743, + 0.0024374357890337706, + 0.0035226689651608467, + 0.005090611986815929, + 0.0035640192218124866, + 0.0024983284529298544, + 0.01804145611822605, + 0.011351943016052246, + 0.008356576785445213, + 0.009344259276986122, + 0.02503211982548237, + 0.01234128326177597, + 0.01600782573223114, + 0.01164510939270258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.029813826084136963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from behind the rider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider an elderly, balding African-American woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse chestnut colored with a flowing black mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rider wear blue jeans, brown leather boots, and a white cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move smoothly while the pair travels along a wide dirt path in a park?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903411775827408 + ], + "opposite_logprob_tokens": [ + -2.464034080505371 + ], + "logprob_reward": -0.08903411775827408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903411775827408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.267109751701355, + 0.06847871094942093, + 0.13598854839801788, + 0.09628155082464218, + 0.09443937987089157, + 0.10001320391893387, + 0.040774233639240265, + 0.05858096107840538, + 0.0633724182844162, + 0.07496127486228943 + ], + "credit_map": [ + 0.05542479828000069, + 0.04292890802025795, + 0.05679827928543091, + 0.02093413658440113, + 0.01802743971347809, + 0.0414980985224247, + 0.015025812201201916, + 0.016472289338707924, + 0.009549763984978199, + 0.009200909174978733, + 0.015359540469944477, + 0.007714735809713602, + 0.00465875118970871, + 0.008011131547391415, + 0.007670896593481302, + 0.006312983576208353, + 0.02928396314382553, + 0.01717687025666237, + 0.029000267386436462, + 0.013497580774128437, + 0.013307321816682816, + 0.01594555750489235, + 0.009211895987391472, + 0.008565079420804977, + 0.010943511500954628, + 0.01218446809798479, + 0.011912412010133266, + 0.008796815760433674, + 0.020403996109962463, + 0.011997741647064686, + 0.01161029189825058, + 0.008432305417954922, + 0.024141691625118256, + 0.011595802381634712, + 0.01648552529513836, + 0.006067720707505941, + 0.006791884079575539, + 0.01605219393968582, + 0.006734787952154875, + 0.006569768767803907, + 0.025314124301075935, + 0.01480162050575018, + 0.0134285818785429, + 0.005858512595295906, + 0.006218298338353634, + 0.020011696964502335, + 0.008071924559772015, + 0.006308439653366804, + 0.006161107216030359, + 0.007931944914162159, + 0.00471752043813467, + 0.0044663697481155396, + 0.00540373008698225, + 0.004714451730251312, + 0.00378896645270288, + 0.0035901435185223818, + 0.016260063275694847, + 0.0068422709591686726, + 0.007584080100059509, + 0.003588181920349598, + 0.004801538307219744, + 0.009974787943065166, + 0.005174131132662296, + 0.004355906508862972, + 0.010070166550576687, + 0.006507952697575092, + 0.010584943927824497, + 0.004834117833524942, + 0.00729299895465374, + 0.010733322240412235, + 0.007850293070077896, + 0.005498616956174374, + 0.015062853693962097, + 0.008399611338973045, + 0.009635637514293194, + 0.006080531515181065, + 0.007247074507176876, + 0.013634074479341507, + 0.005951274652034044, + 0.008950215764343739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.08903411775827408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_tensor([2593, 2593])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_tensor([2593, 2593])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bad9c8259f4554fb02f6331c317fb84f18470329 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_tensor([2593, 2593])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth drone shot, captured at a high frame rate, captures a scene both tranquil and vibrant. Like meticulously crafted emerald trays, the giant leaves of the Amazonian water lily float on the crystal-clear lake. Their edges curl slightly upward, their surfaces covered in delicate radiating veins, their waxy sheen gleaming in the soft afternoon sunlight. As the breeze sweeps across the lake, transparent ripples ripple across the surface, and the giant leaves follow in a gentle, rhythmic rise and fall. Every subtle movement is captured vividly, the surface undulating with a silky, delicate texture. Sunlight dances on the ripples and water droplets on the leaves, creating flowing, shimmering golden highlights. The scene, dominated by emerald green and deep lake blue, creates an idyllic atmosphere of profound tranquility, healing, and vitality.", + "seed": 24177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_2593_seed_24177_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06420844458043576, + "vlm_logprob": -0.06420844458043576, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06420844458043576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2738338112831116 + ], + "opposite_logprob_tokens": [ + -1.5238337516784668 + ], + "logprob_reward": -0.2738338112831116, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2738338112831116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12246468663215637, + 0.0950101688504219, + 0.056546468287706375, + 0.04480404034256935, + 0.08957932889461517, + 0.21666890382766724, + 0.024869773536920547, + 0.04516810178756714, + 0.07106094062328339, + 0.2338276356458664 + ], + "credit_map": [ + 0.016301149502396584, + 0.019965549930930138, + 0.026200270280241966, + 0.01454326044768095, + 0.010835147462785244, + 0.01193027663975954, + 0.011264506727457047, + 0.01142451073974371, + 0.019513526931405067, + 0.006661951541900635, + 0.03735407069325447, + 0.006462638266384602, + 0.00553580978885293, + 0.007290172390639782, + 0.0059256479144096375, + 0.006266341079026461, + 0.00761391082778573, + 0.007969770580530167, + 0.014972718432545662, + 0.004106310661882162, + 0.004133908543735743, + 0.0032198005355894566, + 0.011330291628837585, + 0.0031997524201869965, + 0.008685944601893425, + 0.009353742934763432, + 0.0061522419564425945, + 0.004374369978904724, + 0.003934341948479414, + 0.004709653090685606, + 0.004659085534512997, + 0.0029346547089517117, + 0.021708333864808083, + 0.03971581906080246, + 0.007473081815987825, + 0.003330330364406109, + 0.004810740239918232, + 0.0035678609274327755, + 0.006046554073691368, + 0.0029265936464071274, + 0.04768671840429306, + 0.08780952543020248, + 0.03527040407061577, + 0.008599395863711834, + 0.01117890328168869, + 0.01174230594187975, + 0.005257664248347282, + 0.009123960509896278, + 0.0038076697383075953, + 0.004382008221000433, + 0.0024689456913620234, + 0.0025086586829274893, + 0.002883769106119871, + 0.0033819028176367283, + 0.0033069609198719263, + 0.0021298578940331936, + 0.01556131336838007, + 0.011975608766078949, + 0.003200867446139455, + 0.0023953751660883427, + 0.003305843099951744, + 0.003569443244487047, + 0.0023896307684481144, + 0.0027700155042111874, + 0.01345120556652546, + 0.014092105440795422, + 0.006189474370330572, + 0.005569235887378454, + 0.008985043503344059, + 0.011565834283828735, + 0.005839207675307989, + 0.005368825513869524, + 0.05528536066412926, + 0.10402905941009521, + 0.009965277276933193, + 0.008794700726866722, + 0.012662927620112896, + 0.009397507645189762, + 0.02409696951508522, + 0.009595797397196293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.2738338112831116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016095347702503204 + ], + "opposite_logprob_tokens": [ + -4.141095161437988 + ], + "logprob_reward": -0.016095347702503204, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016095347702503204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16355577111244202, + 0.15786099433898926, + 0.08310653269290924, + 0.035295698791742325, + 0.03070051409304142, + 0.26082664728164673, + 0.06346721202135086, + 0.04148320481181145, + 0.04809993505477905, + 0.11560353636741638 + ], + "credit_map": [ + 0.031102387234568596, + 0.020038675516843796, + 0.03896915540099144, + 0.016603024676442146, + 0.011119544506072998, + 0.014892417006194592, + 0.013744550757110119, + 0.01708601601421833, + 0.025610441341996193, + 0.02677568793296814, + 0.029817398637533188, + 0.018795879557728767, + 0.01506662555038929, + 0.014620205387473106, + 0.016790663823485374, + 0.010384097695350647, + 0.014593164436519146, + 0.019442850723862648, + 0.00863843783736229, + 0.011746636591851711, + 0.007635807618498802, + 0.006218722555786371, + 0.009468593634665012, + 0.0053623151034116745, + 0.0058136554434895515, + 0.005286173429340124, + 0.005051346495747566, + 0.0029869198333472013, + 0.0037603576201945543, + 0.004094548523426056, + 0.00456407992169261, + 0.003738617990165949, + 0.0042139096185564995, + 0.004246614873409271, + 0.008685996755957603, + 0.0026316214352846146, + 0.0029649948701262474, + 0.00363696226850152, + 0.002356672426685691, + 0.00196374231018126, + 0.05397091060876846, + 0.10607236623764038, + 0.04014992713928223, + 0.01007042359560728, + 0.017102137207984924, + 0.01435945089906454, + 0.00824548676609993, + 0.010855956003069878, + 0.016082879155874252, + 0.018821969628334045, + 0.0038163086865097284, + 0.0034603450912982225, + 0.007064759731292725, + 0.00677098985761404, + 0.0032012679148465395, + 0.0042486912570893764, + 0.007783427834510803, + 0.0066338698379695415, + 0.004730800166726112, + 0.004421893507242203, + 0.005202732514590025, + 0.0053633819334208965, + 0.003695498453453183, + 0.00365160102955997, + 0.009873460046947002, + 0.010882936418056488, + 0.004141661338508129, + 0.004110234789550304, + 0.005645215976983309, + 0.006166994571685791, + 0.0039974781684577465, + 0.0032819570042192936, + 0.04240752384066582, + 0.017286524176597595, + 0.011914895847439766, + 0.00777721032500267, + 0.007723035756498575, + 0.007322580553591251, + 0.013947028666734695, + 0.007224735338240862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.016095347702503204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554659508168697 + ], + "opposite_logprob_tokens": [ + -4.38755464553833 + ], + "logprob_reward": -0.012554659508168697, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554659508168697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835050135850906, + 0.06896717101335526, + 0.14653320610523224, + 0.027935102581977844, + 0.12689049541950226, + 0.2816186249256134, + 0.1190778985619545, + 0.021849289536476135, + 0.03997994214296341, + 0.09879779815673828 + ], + "credit_map": [ + 0.01164606586098671, + 0.010810750536620617, + 0.00819410104304552, + 0.007775194942951202, + 0.005402622744441032, + 0.008356374688446522, + 0.0073000406846404076, + 0.008865352720022202, + 0.008410644717514515, + 0.007242647465318441, + 0.01578216254711151, + 0.007209693547338247, + 0.006723362486809492, + 0.008210713975131512, + 0.009679296985268593, + 0.0057086520828306675, + 0.027003305032849312, + 0.03398137167096138, + 0.017116857692599297, + 0.01657838374376297, + 0.01272137276828289, + 0.011237400583922863, + 0.017815669998526573, + 0.010078858584165573, + 0.005095918662846088, + 0.00540007883682847, + 0.0038145275320857763, + 0.0024628781247884035, + 0.0026224537286907434, + 0.002911192364990711, + 0.002842394169420004, + 0.0027856568340212107, + 0.020336205139756203, + 0.049291592091321945, + 0.031262509524822235, + 0.006693908479064703, + 0.0068837497383356094, + 0.005309301428496838, + 0.003415132174268365, + 0.0036980812437832355, + 0.056638188660144806, + 0.11440620571374893, + 0.049723055213689804, + 0.01038502249866724, + 0.016552606597542763, + 0.015195930376648903, + 0.007563907187432051, + 0.011153709143400192, + 0.01941613107919693, + 0.025268539786338806, + 0.011013039387762547, + 0.00950993038713932, + 0.012429843656718731, + 0.01607503369450569, + 0.012442619539797306, + 0.012922766618430614, + 0.0044386256486177444, + 0.006526333279907703, + 0.0012057540006935596, + 0.0012237117625772953, + 0.0020627127960324287, + 0.0028934471774846315, + 0.002032308839261532, + 0.0014663946349173784, + 0.009096386842429638, + 0.009855479001998901, + 0.00317327375523746, + 0.0031266706064343452, + 0.004291924182325602, + 0.00462400633841753, + 0.0032986148726195097, + 0.0025135886389762163, + 0.03308480978012085, + 0.026814939454197884, + 0.009388322941958904, + 0.0055365366861224174, + 0.007701271679252386, + 0.006120679434388876, + 0.005648966412991285, + 0.004502275492995977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.012554659508168697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068601980805397 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068601980805397, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068601980805397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2390749603509903, + 0.07812013477087021, + 0.11687914282083511, + 0.11065835505723953, + 0.04174210503697395, + 0.21281945705413818, + 0.0735115110874176, + 0.021259207278490067, + 0.02624785713851452, + 0.0796872153878212 + ], + "credit_map": [ + 0.05520761385560036, + 0.02440689504146576, + 0.07180245965719223, + 0.017884518951177597, + 0.01562504842877388, + 0.019391460344195366, + 0.01644597202539444, + 0.018311021849513054, + 0.00887410156428814, + 0.006815613247454166, + 0.03207622841000557, + 0.0061253891326487064, + 0.004258797504007816, + 0.00606452627107501, + 0.008523953147232533, + 0.005381525028496981, + 0.016046376898884773, + 0.016512826085090637, + 0.03038395196199417, + 0.010580314323306084, + 0.0087817944586277, + 0.009111423045396805, + 0.01964518241584301, + 0.00581728108227253, + 0.015480892732739449, + 0.020467981696128845, + 0.02612883225083351, + 0.00960227195173502, + 0.008721085265278816, + 0.010638842359185219, + 0.008509645238518715, + 0.011108805425465107, + 0.00805168878287077, + 0.006366303656250238, + 0.010930037125945091, + 0.00271071819588542, + 0.003150345291942358, + 0.0031526407692581415, + 0.0048796930350363255, + 0.0025006812065839767, + 0.044598739594221115, + 0.08574056625366211, + 0.039724886417388916, + 0.008642868138849735, + 0.009929428808391094, + 0.010810811072587967, + 0.004790408071130514, + 0.008581765927374363, + 0.012623543851077557, + 0.01846522092819214, + 0.006798231042921543, + 0.005873736459761858, + 0.006106565706431866, + 0.007954058237373829, + 0.0069167097099125385, + 0.008773453533649445, + 0.005168652627617121, + 0.005434488877654076, + 0.0014268826926127076, + 0.0015790399629622698, + 0.0022420757450163364, + 0.0025626560673117638, + 0.0015868486370891333, + 0.0012585640652105212, + 0.00547682074829936, + 0.0067299166694283485, + 0.002070083748549223, + 0.0023616452235728502, + 0.0030217694584280252, + 0.002689890330657363, + 0.002033643424510956, + 0.0018640885828062892, + 0.029003622010350227, + 0.0328906886279583, + 0.003976450301706791, + 0.002869928488507867, + 0.0038050764705985785, + 0.0028447983786463737, + 0.002469231840223074, + 0.001827426254749298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.016068601980805397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024898024275898933 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024898024275898933, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024898024275898933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09569136798381805, + 0.1368381381034851, + 0.08364392817020416, + 0.11030379682779312, + 0.16315241158008575, + 0.16699837148189545, + 0.06649164855480194, + 0.03898291289806366, + 0.08357793092727661, + 0.05431949347257614 + ], + "credit_map": [ + 0.01502674538642168, + 0.01714370585978031, + 0.007885872386395931, + 0.01379917748272419, + 0.006517405156046152, + 0.009624582715332508, + 0.013331964612007141, + 0.012361910194158554, + 0.024030981585383415, + 0.018274547532200813, + 0.04532549902796745, + 0.011202817782759666, + 0.009647614322602749, + 0.008797889575362206, + 0.010057478211820126, + 0.009501309134066105, + 0.016730215400457382, + 0.019912566989660263, + 0.006642199587076902, + 0.008400468155741692, + 0.007537578698247671, + 0.006781252566725016, + 0.01126598846167326, + 0.006373660173267126, + 0.01791115663945675, + 0.016721060499548912, + 0.02705923654139042, + 0.009203829802572727, + 0.010482214391231537, + 0.009713172912597656, + 0.010123246349394321, + 0.009089869447052479, + 0.027254067361354828, + 0.0647106021642685, + 0.03650638461112976, + 0.008316454477608204, + 0.008761636912822723, + 0.006594350095838308, + 0.0047050644643604755, + 0.006303861271589994, + 0.03472795709967613, + 0.06727687269449234, + 0.029718240723013878, + 0.006500690709799528, + 0.009099074639379978, + 0.0087754987180233, + 0.004352669697254896, + 0.006547378376126289, + 0.013503052294254303, + 0.013855716213583946, + 0.0056328740902245045, + 0.00485644955188036, + 0.00971157941967249, + 0.007384789176285267, + 0.005573154892772436, + 0.005974035710096359, + 0.011729774065315723, + 0.013722746632993221, + 0.001885297941043973, + 0.0015540807507932186, + 0.0032847237307578325, + 0.0023456537164747715, + 0.0020182952284812927, + 0.0024423415306955576, + 0.018329529091715813, + 0.017790518701076508, + 0.0068357037380337715, + 0.0064155831933021545, + 0.009894528426229954, + 0.01154792495071888, + 0.006963347550481558, + 0.005800788290798664, + 0.01458148006349802, + 0.009068685583770275, + 0.004948913585394621, + 0.003550361143425107, + 0.00485730217769742, + 0.0032102905679494143, + 0.009357143193483353, + 0.004745318088680506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.0024898024275898933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth drone shot, captured at a high frame rate, captures a scene both tranquil and vibrant. Like meticulously crafted emerald trays, the giant leaves of the Amazonian water lily float on the crystal-clear lake. Their edges curl slightly upward, their surfaces covered in delicate radiating veins, their waxy sheen gleaming in the soft afternoon sunlight. As the breeze sweeps across the lake, transparent ripples ripple across the surface, and the giant leaves follow in a gentle, rhythmic rise and fall. Every subtle movement is captured vividly, the surface undulating with a silky, delicate texture. Sunlight dances on the ripples and water droplets on the leaves, creating flowing, shimmering golden highlights. The scene, dominated by emerald green and deep lake blue, creates an idyllic atmosphere of profound tranquility, healing, and vitality.", + "seed": 24178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_2593_seed_24178_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05418567592278123, + "vlm_logprob": -0.05418567592278123, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.05418567592278123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23736678063869476 + ], + "opposite_logprob_tokens": [ + -1.612366795539856 + ], + "logprob_reward": -0.23736678063869476, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23736678063869476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1801135390996933, + 0.043874017894268036, + 0.12604935467243195, + 0.10624664276838303, + 0.07089538127183914, + 0.09976574778556824, + 0.07113838195800781, + 0.09157335013151169, + 0.13566645979881287, + 0.07467717677354813 + ], + "credit_map": [ + 0.04506982862949371, + 0.01514242310076952, + 0.021283447742462158, + 0.019278258085250854, + 0.016949445009231567, + 0.013437624089419842, + 0.01959390379488468, + 0.029358593747019768, + 0.011219079606235027, + 0.004092532675713301, + 0.00410070177167654, + 0.0049334922805428505, + 0.0043204049579799175, + 0.003925520461052656, + 0.006341979838907719, + 0.004940306767821312, + 0.025521261617541313, + 0.011937549337744713, + 0.015601723454892635, + 0.01164229679852724, + 0.011103176511824131, + 0.013180806301534176, + 0.013318780809640884, + 0.023743748664855957, + 0.01550765335559845, + 0.008186009712517262, + 0.013758335262537003, + 0.012916691601276398, + 0.010320263914763927, + 0.010957195423543453, + 0.01411275751888752, + 0.020487742498517036, + 0.009596839547157288, + 0.005967630539089441, + 0.009820302948355675, + 0.008128887973725796, + 0.00888845231384039, + 0.006409741938114166, + 0.012111173011362553, + 0.009972354397177696, + 0.014896884560585022, + 0.00610541133210063, + 0.023048505187034607, + 0.009055888280272484, + 0.006849281024187803, + 0.008016034960746765, + 0.018690481781959534, + 0.013103251345455647, + 0.008500846102833748, + 0.010821535252034664, + 0.011828694492578506, + 0.0072402264922857285, + 0.00967856403440237, + 0.006710391957312822, + 0.009666423313319683, + 0.006691703572869301, + 0.009507772512733936, + 0.010122125968337059, + 0.01441982388496399, + 0.008677000179886818, + 0.01874409057199955, + 0.010995295830070972, + 0.010122242383658886, + 0.008984996937215328, + 0.01836283877491951, + 0.01090297382324934, + 0.021136056631803513, + 0.015081500634551048, + 0.022653764113783836, + 0.014607398770749569, + 0.0184207521378994, + 0.014501168392598629, + 0.011170553974807262, + 0.006435975898057222, + 0.012126026675105095, + 0.009093658067286015, + 0.012212087400257587, + 0.007142760790884495, + 0.010609500110149384, + 0.005886610597372055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.23736678063869476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00465867854654789 + ], + "opposite_logprob_tokens": [ + -5.3796586990356445 + ], + "logprob_reward": -0.00465867854654789, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00465867854654789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1356765478849411, + 0.048331473022699356, + 0.14174143970012665, + 0.10022035986185074, + 0.09176354110240936, + 0.07581303268671036, + 0.07754956930875778, + 0.0251625869423151, + 0.1955130249261856, + 0.1082284152507782 + ], + "credit_map": [ + 0.03534923493862152, + 0.009015209972858429, + 0.015845242887735367, + 0.012727271765470505, + 0.015309148468077183, + 0.00928453542292118, + 0.019189435988664627, + 0.01895647495985031, + 0.009198568761348724, + 0.004368302877992392, + 0.006352773867547512, + 0.006611356977373362, + 0.005623433738946915, + 0.00453987019136548, + 0.006452628877013922, + 0.005184540059417486, + 0.021549571305513382, + 0.012750371359288692, + 0.01845775917172432, + 0.011377313174307346, + 0.011279181577265263, + 0.0196375735104084, + 0.01110521238297224, + 0.035584449768066406, + 0.015959421172738075, + 0.009545194916427135, + 0.010003626346588135, + 0.01303517259657383, + 0.014886575751006603, + 0.011294838041067123, + 0.01504835020750761, + 0.01044718362390995, + 0.01044753473252058, + 0.00949979480355978, + 0.009030255489051342, + 0.012455458752810955, + 0.010047852993011475, + 0.01631864905357361, + 0.015168989077210426, + 0.008795010857284069, + 0.009377258829772472, + 0.006568214390426874, + 0.010827271267771721, + 0.008178193122148514, + 0.011198370717465878, + 0.007681393064558506, + 0.015438727103173733, + 0.0065436032600700855, + 0.009008028544485569, + 0.00839057844132185, + 0.011007791385054588, + 0.006602899171411991, + 0.014697321690618992, + 0.010175298899412155, + 0.007889380678534508, + 0.00977826863527298, + 0.0031999971251934767, + 0.002863973844796419, + 0.0030120229348540306, + 0.0027424658183008432, + 0.0038232498336583376, + 0.00371616636402905, + 0.00305178458802402, + 0.002752928528934717, + 0.028054531663656235, + 0.014652171172201633, + 0.04352574050426483, + 0.01909611001610756, + 0.024560686200857162, + 0.019521690905094147, + 0.025767402723431587, + 0.020334681496024132, + 0.015553882345557213, + 0.011603877879679203, + 0.017242539674043655, + 0.012693274766206741, + 0.017974315211176872, + 0.009648609906435013, + 0.015161469578742981, + 0.008350437507033348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.00465867854654789 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001046348363161087 + ], + "opposite_logprob_tokens": [ + -6.876046180725098 + ], + "logprob_reward": -0.001046348363161087, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001046348363161087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19923029839992523, + 0.15608493983745575, + 0.08901268243789673, + 0.06405413150787354, + 0.06392394751310349, + 0.037423353642225266, + 0.0596088282763958, + 0.04583938047289848, + 0.2260308861732483, + 0.058791544288396835 + ], + "credit_map": [ + 0.04666614159941673, + 0.018604231998324394, + 0.019851837307214737, + 0.021377408877015114, + 0.021627813577651978, + 0.016452860087156296, + 0.024896686896681786, + 0.029753316193819046, + 0.035965707153081894, + 0.011980706825852394, + 0.01416033785790205, + 0.016914306208491325, + 0.014520146884024143, + 0.01320553570985794, + 0.019335906952619553, + 0.030002297833561897, + 0.016190292313694954, + 0.008293360471725464, + 0.011799764819443226, + 0.008239329792559147, + 0.008998865261673927, + 0.009254705160856247, + 0.00877771619707346, + 0.01745864935219288, + 0.009565240703523159, + 0.00675706984475255, + 0.007585955783724785, + 0.0093305092304945, + 0.008005212061107159, + 0.006487708073109388, + 0.007895488291978836, + 0.008426941931247711, + 0.007388944271951914, + 0.007898781448602676, + 0.0067052277736365795, + 0.009315119124948978, + 0.006467039696872234, + 0.007887322455644608, + 0.012761637568473816, + 0.005499879829585552, + 0.004484525416046381, + 0.0031274633947759867, + 0.00693935202434659, + 0.004659547004848719, + 0.0054697562009096146, + 0.0029994603246450424, + 0.005715209525078535, + 0.004028039518743753, + 0.006588672287762165, + 0.0065658846870064735, + 0.011440270580351353, + 0.00747219892218709, + 0.008404797874391079, + 0.00478727463632822, + 0.0075097340159118176, + 0.006839994341135025, + 0.0053543285466730595, + 0.006536030676215887, + 0.00955141894519329, + 0.004096442833542824, + 0.006384249310940504, + 0.004526285920292139, + 0.0046109165996313095, + 0.004779705312103033, + 0.03231070563197136, + 0.017105450853705406, + 0.0510435551404953, + 0.022610878571867943, + 0.0261888075619936, + 0.02281167171895504, + 0.029955383390188217, + 0.024004435166716576, + 0.008770914748311043, + 0.0053244782611727715, + 0.00977984257042408, + 0.007073518820106983, + 0.009831279516220093, + 0.005777215119451284, + 0.007752572186291218, + 0.004481722135096788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.001046348363161087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343639940023422 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.026343639940023422, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343639940023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253614217042923, + 0.1433354616165161, + 0.18116052448749542, + 0.04962297901511192, + 0.105470210313797, + 0.10897690057754517, + 0.04142439737915993, + 0.03761131688952446, + 0.03286188840866089, + 0.07417488843202591 + ], + "credit_map": [ + 0.05921851471066475, + 0.01752440258860588, + 0.028590640053153038, + 0.02249046601355076, + 0.018406648188829422, + 0.01641613431274891, + 0.02647220343351364, + 0.03624238818883896, + 0.03255781903862953, + 0.010947896167635918, + 0.014497311785817146, + 0.016873814165592194, + 0.014110902324318886, + 0.012591637670993805, + 0.013088681735098362, + 0.028667397797107697, + 0.030231870710849762, + 0.015692656859755516, + 0.023921070620417595, + 0.013896103017032146, + 0.01413534115999937, + 0.02069232054054737, + 0.017217082902789116, + 0.045374080538749695, + 0.0059768739156425, + 0.0049246931448578835, + 0.006926397327333689, + 0.006016919855028391, + 0.006295888684689999, + 0.004667582456022501, + 0.008171633817255497, + 0.00664298003539443, + 0.014792186208069324, + 0.010442177765071392, + 0.02877948246896267, + 0.008529291488230228, + 0.008701719343662262, + 0.011546439491212368, + 0.01073111966252327, + 0.011947788298130035, + 0.01465562917292118, + 0.00804679561406374, + 0.024806194007396698, + 0.011339470744132996, + 0.007554061245173216, + 0.01006572786718607, + 0.019134968519210815, + 0.013374038971960545, + 0.004451059270650148, + 0.005156129598617554, + 0.006521566770970821, + 0.004225737880915403, + 0.0066140443086624146, + 0.0036660057958215475, + 0.005609329789876938, + 0.005180522799491882, + 0.0045309788547456264, + 0.0038088758010417223, + 0.0062751565128564835, + 0.004446717444807291, + 0.006128883920609951, + 0.003900539595633745, + 0.004699219949543476, + 0.003820944344624877, + 0.005242261104285717, + 0.00224553095176816, + 0.006155220326036215, + 0.003818974131718278, + 0.004290238954126835, + 0.0027548130601644516, + 0.005241571459919214, + 0.003113276557996869, + 0.011314606294035912, + 0.0064635006710886955, + 0.012786977924406528, + 0.00861850194633007, + 0.011850710958242416, + 0.0070261661894619465, + 0.010120577178895473, + 0.005993844009935856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026343639940023422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001512932125478983 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001512932125478983, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001512932125478983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08833523094654083, + 0.03603080287575722, + 0.06076808273792267, + 0.050453513860702515, + 0.05621813237667084, + 0.042746495455503464, + 0.10079123824834824, + 0.04812072589993477, + 0.4647713303565979, + 0.05176438391208649 + ], + "credit_map": [ + 0.017643267288804054, + 0.007222411222755909, + 0.011029358021914959, + 0.0100974440574646, + 0.007889516651630402, + 0.0071840714663267136, + 0.016866587102413177, + 0.010402579791843891, + 0.00855225883424282, + 0.0033842152915894985, + 0.003523891093209386, + 0.003229694440960884, + 0.003794913413003087, + 0.002977722091600299, + 0.004001460038125515, + 0.0065666488371789455, + 0.010559833608567715, + 0.005383649840950966, + 0.0076857940293848515, + 0.006342582404613495, + 0.007015769835561514, + 0.005708628334105015, + 0.007413526065647602, + 0.010658306069672108, + 0.006933839060366154, + 0.004165433347225189, + 0.0058304620906710625, + 0.007559949532151222, + 0.006955656222999096, + 0.0049437168054282665, + 0.007210039999336004, + 0.006854421459138393, + 0.006630839314311743, + 0.00544179929420352, + 0.006685055326670408, + 0.007576644886285067, + 0.007489378564059734, + 0.0068198139779269695, + 0.00892696063965559, + 0.006647649686783552, + 0.005300029646605253, + 0.0034476809669286013, + 0.009423960000276566, + 0.0047489432618021965, + 0.006154037546366453, + 0.0037926724180579185, + 0.0049653565511107445, + 0.004913819953799248, + 0.011719921603798866, + 0.00896022841334343, + 0.02000896818935871, + 0.009247472509741783, + 0.016486257314682007, + 0.009353132918477058, + 0.009203657507896423, + 0.015811609104275703, + 0.005434910301119089, + 0.0072577884420752525, + 0.007698222529143095, + 0.005810518749058247, + 0.008160693570971489, + 0.004950774367898703, + 0.005314949434250593, + 0.003492873627692461, + 0.06516418606042862, + 0.03516879305243492, + 0.10445970296859741, + 0.04598125070333481, + 0.05697856843471527, + 0.04684210196137428, + 0.06157050281763077, + 0.04860629141330719, + 0.007407736498862505, + 0.0048975152894854546, + 0.0088859423995018, + 0.006310952827334404, + 0.008249646052718163, + 0.004715634509921074, + 0.007248351816087961, + 0.004048611968755722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.001512932125478983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth drone shot, captured at a high frame rate, captures a scene both tranquil and vibrant. Like meticulously crafted emerald trays, the giant leaves of the Amazonian water lily float on the crystal-clear lake. Their edges curl slightly upward, their surfaces covered in delicate radiating veins, their waxy sheen gleaming in the soft afternoon sunlight. As the breeze sweeps across the lake, transparent ripples ripple across the surface, and the giant leaves follow in a gentle, rhythmic rise and fall. Every subtle movement is captured vividly, the surface undulating with a silky, delicate texture. Sunlight dances on the ripples and water droplets on the leaves, creating flowing, shimmering golden highlights. The scene, dominated by emerald green and deep lake blue, creates an idyllic atmosphere of profound tranquility, healing, and vitality.", + "seed": 24179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_2593_seed_24179_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027231045649386942, + "vlm_logprob": -0.027231045649386942, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.027231045649386942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1144048273563385 + ], + "opposite_logprob_tokens": [ + -2.3644049167633057 + ], + "logprob_reward": -0.1144048273563385, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1144048273563385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4878152012825012, + 0.05696431174874306, + 0.12030160427093506, + 0.13473235070705414, + 0.062174241989851, + 0.03245871141552925, + 0.03903927281498909, + 0.019529569894075394, + 0.01897086389362812, + 0.02801385708153248 + ], + "credit_map": [ + 0.02013518288731575, + 0.02122645452618599, + 0.043803147971630096, + 0.14107631146907806, + 0.18001830577850342, + 0.04122638329863548, + 0.025638164952397346, + 0.01469123549759388, + 0.005832114722579718, + 0.007817610166966915, + 0.006680067628622055, + 0.007469904609024525, + 0.006654552649706602, + 0.007185003254562616, + 0.008622459135949612, + 0.00670260563492775, + 0.01507054828107357, + 0.011221691966056824, + 0.0182101558893919, + 0.014601502567529678, + 0.019529134035110474, + 0.014922073110938072, + 0.013566438108682632, + 0.013180074281990528, + 0.013292496092617512, + 0.010544408112764359, + 0.010110493749380112, + 0.01337121520191431, + 0.037652768194675446, + 0.01771966740489006, + 0.011229036375880241, + 0.020812273025512695, + 0.005840822122991085, + 0.005036538932472467, + 0.006229748949408531, + 0.008223582990467548, + 0.010579249821603298, + 0.011658594943583012, + 0.006721964571624994, + 0.007883741520345211, + 0.004335631150752306, + 0.002999385818839073, + 0.003805497894063592, + 0.004892209079116583, + 0.0048104580491781235, + 0.0025293738581240177, + 0.0045203943736851215, + 0.004565765615552664, + 0.004047285299748182, + 0.0037443917244672775, + 0.004007421899586916, + 0.0037143281660974026, + 0.008794455789029598, + 0.005950169172137976, + 0.003751030657440424, + 0.005030193366110325, + 0.0025205097626894712, + 0.0017778007313609123, + 0.002053684089332819, + 0.0019437525188550353, + 0.0032732277177274227, + 0.0037531007546931505, + 0.0019350402290001512, + 0.0022724561858922243, + 0.0018589119426906109, + 0.0013716993853449821, + 0.0018320129020139575, + 0.00205943058244884, + 0.002593566896393895, + 0.0024118905421346426, + 0.0033604747150093317, + 0.0034828772768378258, + 0.002147822408005595, + 0.0029082426335662603, + 0.00198728172108531, + 0.002267087809741497, + 0.004512575455009937, + 0.0067613390274345875, + 0.0031997812911868095, + 0.004229729995131493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.1144048273563385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017296605510637164 + ], + "opposite_logprob_tokens": [ + -6.376729488372803 + ], + "logprob_reward": -0.0017296605510637164, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017296605510637164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6503640413284302, + 0.04770291969180107, + 0.08079725503921509, + 0.06824224442243576, + 0.014010533690452576, + 0.018899215385317802, + 0.027564235031604767, + 0.014061588793992996, + 0.012368708848953247, + 0.06598930060863495 + ], + "credit_map": [ + 0.023127546533942223, + 0.02522061951458454, + 0.05949313938617706, + 0.18993671238422394, + 0.24805937707424164, + 0.0545242503285408, + 0.030487561598420143, + 0.01951480284333229, + 0.004965758882462978, + 0.006577306427061558, + 0.005423352122306824, + 0.004253399558365345, + 0.007312803063541651, + 0.006503199692815542, + 0.007297605741769075, + 0.005369492340832949, + 0.0068346536718308926, + 0.00658261775970459, + 0.011067076586186886, + 0.007284512277692556, + 0.021171091124415398, + 0.011684800498187542, + 0.008949763141572475, + 0.007222736719995737, + 0.006000480614602566, + 0.004304691683501005, + 0.005501622334122658, + 0.007572449278086424, + 0.01785212755203247, + 0.012953657656908035, + 0.004409490618854761, + 0.009647720493376255, + 0.0011099997209385037, + 0.0012642398942261934, + 0.001217043143697083, + 0.0012251859297975898, + 0.0031143680680543184, + 0.0019878942985087633, + 0.0019595951307564974, + 0.002132207155227661, + 0.001763541135005653, + 0.0013229547766968608, + 0.0017565469024702907, + 0.0015034846728667617, + 0.0034398199059069157, + 0.001900251372717321, + 0.0023470257874578238, + 0.004865590482950211, + 0.0030688645783811808, + 0.0035669133067131042, + 0.0031578713096678257, + 0.002457440597936511, + 0.004688011948019266, + 0.003701629349961877, + 0.0027015521191060543, + 0.004221951588988304, + 0.001573264366015792, + 0.001589312800206244, + 0.002073879586532712, + 0.0015718925278633833, + 0.0024866070598363876, + 0.0021313203033059835, + 0.001271478133276105, + 0.0013638335512951016, + 0.0014811415458098054, + 0.0009928703075274825, + 0.0011261558393016458, + 0.0009827270405367017, + 0.0016843025805428624, + 0.001384518458507955, + 0.0017067445442080498, + 0.003010248765349388, + 0.006817637011408806, + 0.007987407967448235, + 0.006428828928619623, + 0.005204085726290941, + 0.008288897573947906, + 0.01058833859860897, + 0.0076992325484752655, + 0.012974870391190052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0017296605510637164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036253698635846376 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.0036253698635846376, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036253698635846376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30020228028297424, + 0.09117616713047028, + 0.12458200007677078, + 0.07082713395357132, + 0.03919931501150131, + 0.10010464489459991, + 0.0490705706179142, + 0.053876202553510666, + 0.05175351724028587, + 0.11920817196369171 + ], + "credit_map": [ + 0.013690968975424767, + 0.017689064145088196, + 0.03293944150209427, + 0.07580778002738953, + 0.08989174664020538, + 0.03408290445804596, + 0.01659243367612362, + 0.019507916644215584, + 0.008613014593720436, + 0.007968411780893803, + 0.014287975616753101, + 0.011431915685534477, + 0.016228821128606796, + 0.012155012227594852, + 0.010762615129351616, + 0.009728392586112022, + 0.013788418844342232, + 0.013000424951314926, + 0.01250795740634203, + 0.01132654957473278, + 0.028319478034973145, + 0.013739576563239098, + 0.014277299866080284, + 0.017622284591197968, + 0.006753040477633476, + 0.006855076178908348, + 0.008060443215072155, + 0.007512995041906834, + 0.011598104611039162, + 0.014772485941648483, + 0.007044884841889143, + 0.008230101317167282, + 0.004942253697663546, + 0.004511239938437939, + 0.003558925585821271, + 0.005277091637253761, + 0.007060416042804718, + 0.004332997836172581, + 0.005539048463106155, + 0.0039773378521203995, + 0.007890868932008743, + 0.004223653115332127, + 0.007132875267416239, + 0.009939239360392094, + 0.02082035131752491, + 0.0073096295818686485, + 0.011826474219560623, + 0.030961547046899796, + 0.00508192228153348, + 0.005394282750785351, + 0.005045650061219931, + 0.004965633619576693, + 0.006704047787934542, + 0.008230248466134071, + 0.006188640370965004, + 0.007460139226168394, + 0.008560548536479473, + 0.007429505232721567, + 0.006504591088742018, + 0.0061487178318202496, + 0.007837604731321335, + 0.005947368685156107, + 0.005929336883127689, + 0.005518526304513216, + 0.00824621133506298, + 0.0069564953446388245, + 0.006387676112353802, + 0.0049880207516252995, + 0.006389009766280651, + 0.004511557053774595, + 0.0049253362230956554, + 0.009349208325147629, + 0.016097567975521088, + 0.016840258613228798, + 0.010020510293543339, + 0.008689955808222294, + 0.016070540994405746, + 0.020142734050750732, + 0.011866871267557144, + 0.019479727372527122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0036253698635846376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196828007698059 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014196828007698059, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196828007698059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10733933001756668, + 0.10141077637672424, + 0.09391815215349197, + 0.26815077662467957, + 0.06350480020046234, + 0.042145196348428726, + 0.05109456181526184, + 0.06721918284893036, + 0.15642234683036804, + 0.04879485443234444 + ], + "credit_map": [ + 0.012603732757270336, + 0.01331720408052206, + 0.013170775026082993, + 0.009749109856784344, + 0.018199225887656212, + 0.01829846389591694, + 0.010530054569244385, + 0.011470768600702286, + 0.00792304053902626, + 0.01267206110060215, + 0.013016766868531704, + 0.00964287668466568, + 0.01926642842590809, + 0.014627768658101559, + 0.013198566623032093, + 0.01106327399611473, + 0.011046948842704296, + 0.012428847141563892, + 0.010479714721441269, + 0.011699297465384007, + 0.015268541872501373, + 0.009322190657258034, + 0.012417148798704147, + 0.011255464516580105, + 0.027944615110754967, + 0.01944125071167946, + 0.019624408334493637, + 0.02850434184074402, + 0.0760081484913826, + 0.03614943102002144, + 0.019203415140509605, + 0.04127519950270653, + 0.004710725042968988, + 0.004723109304904938, + 0.005911173764616251, + 0.008926807902753353, + 0.011026479303836823, + 0.010886565782129765, + 0.008727053180336952, + 0.008592886850237846, + 0.005835187155753374, + 0.0034056196454912424, + 0.00830195751041174, + 0.0064442092552781105, + 0.007190885953605175, + 0.003252664115279913, + 0.003617590991780162, + 0.004097086377441883, + 0.0033535538241267204, + 0.004461184144020081, + 0.004290091339498758, + 0.004007827024906874, + 0.0129522904753685, + 0.005404727533459663, + 0.006843550596386194, + 0.009781340137124062, + 0.006879521068185568, + 0.005505599081516266, + 0.010335123166441917, + 0.009401902556419373, + 0.010548541322350502, + 0.009084952995181084, + 0.007419303059577942, + 0.008044240064918995, + 0.015541264787316322, + 0.020127901807427406, + 0.013255341909825802, + 0.012127800844609737, + 0.027022648602724075, + 0.027819879353046417, + 0.021570174023509026, + 0.018957363441586494, + 0.004783108364790678, + 0.003963574767112732, + 0.004575223661959171, + 0.0033308102283626795, + 0.007589748129248619, + 0.010494539514183998, + 0.005070711486041546, + 0.008987138979136944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.014196828007698059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198542468249798 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198542468249798, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198542468249798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6499330401420593, + 0.04298100620508194, + 0.042921338230371475, + 0.17542997002601624, + 0.013817375525832176, + 0.02349328063428402, + 0.010259553790092468, + 0.012508037500083447, + 0.007913559675216675, + 0.02074282057583332 + ], + "credit_map": [ + 0.021936429664492607, + 0.024932242929935455, + 0.05998915433883667, + 0.1918799728155136, + 0.24559561908245087, + 0.05529581382870674, + 0.03049808368086815, + 0.019805725663900375, + 0.0034912475384771824, + 0.004443672485649586, + 0.005814000032842159, + 0.005171151831746101, + 0.006180268246680498, + 0.005787282716482878, + 0.006641358602792025, + 0.005452023819088936, + 0.0031634983606636524, + 0.003880508942529559, + 0.004755724687129259, + 0.004843666218221188, + 0.008029306307435036, + 0.006070754490792751, + 0.005192344542592764, + 0.006985533516854048, + 0.013983115553855896, + 0.010408120229840279, + 0.012477243319153786, + 0.01652749627828598, + 0.044152818620204926, + 0.039882175624370575, + 0.013419737108051777, + 0.02457926608622074, + 0.001091139973141253, + 0.0011343786027282476, + 0.0012683745007961988, + 0.0017807043623179197, + 0.0024804037529975176, + 0.0026176029350608587, + 0.0015807924792170525, + 0.0018639782210811973, + 0.0015986522193998098, + 0.0009898054413497448, + 0.0017837764462456107, + 0.0022987325210124254, + 0.004931905772536993, + 0.0019515423336997628, + 0.002813836792483926, + 0.007125029340386391, + 0.001050748978741467, + 0.0007752585806883872, + 0.0012943553738296032, + 0.0011874994961544871, + 0.002232381608337164, + 0.001302216318435967, + 0.0011688645463436842, + 0.0012482288293540478, + 0.0013219073880463839, + 0.0011912937043234706, + 0.0010587386786937714, + 0.0013426747173070908, + 0.0024173196870833635, + 0.0020840696524828672, + 0.0015413440996780992, + 0.0015506899217143655, + 0.0007491110591217875, + 0.0008111859788186848, + 0.0009876209078356624, + 0.0009687287383712828, + 0.0012503336183726788, + 0.0009977829176932573, + 0.0008167048799805343, + 0.001332091516815126, + 0.0023294587153941393, + 0.00281819817610085, + 0.002255245577543974, + 0.0020300839096307755, + 0.0037951841950416565, + 0.0028725680895149708, + 0.0023674387484788895, + 0.002274641999974847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002198542468249798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth drone shot, captured at a high frame rate, captures a scene both tranquil and vibrant. Like meticulously crafted emerald trays, the giant leaves of the Amazonian water lily float on the crystal-clear lake. Their edges curl slightly upward, their surfaces covered in delicate radiating veins, their waxy sheen gleaming in the soft afternoon sunlight. As the breeze sweeps across the lake, transparent ripples ripple across the surface, and the giant leaves follow in a gentle, rhythmic rise and fall. Every subtle movement is captured vividly, the surface undulating with a silky, delicate texture. Sunlight dances on the ripples and water droplets on the leaves, creating flowing, shimmering golden highlights. The scene, dominated by emerald green and deep lake blue, creates an idyllic atmosphere of profound tranquility, healing, and vitality.", + "seed": 24180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_15/index_2593_seed_24180_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016672268323600293, + "vlm_logprob": -0.016672268323600293, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.016672268323600293, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.061222899705171585 + ], + "opposite_logprob_tokens": [ + -3.061222791671753 + ], + "logprob_reward": -0.061222899705171585, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.061222899705171585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1763315349817276, + 0.0603189654648304, + 0.03133545070886612, + 0.20624926686286926, + 0.11111485213041306, + 0.07979951053857803, + 0.0999789834022522, + 0.16000981628894806, + 0.03635488823056221, + 0.03850674256682396 + ], + "credit_map": [ + 0.01811893843114376, + 0.01994265615940094, + 0.02427338995039463, + 0.02243150770664215, + 0.023405855521559715, + 0.020308295264840126, + 0.020563999190926552, + 0.027286915108561516, + 0.006110429763793945, + 0.006578600965440273, + 0.009158375672996044, + 0.006531267426908016, + 0.009992057457566261, + 0.006122758146375418, + 0.0060770767740905285, + 0.009748401120305061, + 0.0031005830969661474, + 0.002977346070110798, + 0.0029885738622397184, + 0.003239652141928673, + 0.0047639296390116215, + 0.006659964565187693, + 0.002526558004319668, + 0.005078841932117939, + 0.011057793162763119, + 0.011003276333212852, + 0.00649072602391243, + 0.010140637867152691, + 0.07289668172597885, + 0.02637992426753044, + 0.05791005492210388, + 0.01037017535418272, + 0.006579747423529625, + 0.00951789878308773, + 0.01659998670220375, + 0.007682331372052431, + 0.026255764067173004, + 0.016862748190760612, + 0.02043912373483181, + 0.007177260704338551, + 0.00504258880391717, + 0.004178525414317846, + 0.011238692328333855, + 0.006303697358816862, + 0.028650568798184395, + 0.010001505725085735, + 0.007296382449567318, + 0.007087561301887035, + 0.0076875267550349236, + 0.005193596705794334, + 0.0272507406771183, + 0.0053812856785953045, + 0.03400958329439163, + 0.012650905176997185, + 0.00373101606965065, + 0.004074330907315016, + 0.01680932007730007, + 0.0088187325745821, + 0.01846105419099331, + 0.01090757641941309, + 0.022464649751782417, + 0.01836196519434452, + 0.021611928939819336, + 0.042574603110551834, + 0.003870979882776737, + 0.008678868412971497, + 0.0032808990217745304, + 0.005246346350759268, + 0.0038453093729913235, + 0.0043600392527878284, + 0.003220499726012349, + 0.003851949470117688, + 0.007215591613203287, + 0.006090392824262381, + 0.0033071935176849365, + 0.003805463435128331, + 0.004701411817222834, + 0.005254932679235935, + 0.002976702991873026, + 0.005155058111995459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.061222899705171585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028329261112958193 + ], + "opposite_logprob_tokens": [ + -5.877832889556885 + ], + "logprob_reward": -0.0028329261112958193, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028329261112958193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19103071093559265, + 0.07607746124267578, + 0.10678333789110184, + 0.3288969099521637, + 0.03416631370782852, + 0.046576257795095444, + 0.06746488809585571, + 0.09001250565052032, + 0.02099395915865898, + 0.037997595965862274 + ], + "credit_map": [ + 0.027359738945961, + 0.024594739079475403, + 0.017966056242585182, + 0.021854940801858902, + 0.03228519484400749, + 0.03063538856804371, + 0.018475603312253952, + 0.017859047278761864, + 0.011448407545685768, + 0.0169980451464653, + 0.006261084228754044, + 0.0072338473983109, + 0.010370476171374321, + 0.007591498550027609, + 0.0059684328734874725, + 0.01020567025989294, + 0.010561924427747726, + 0.013259230181574821, + 0.011194542050361633, + 0.012432031333446503, + 0.014940683729946613, + 0.024690857157111168, + 0.008935417048633099, + 0.010768651962280273, + 0.012202375568449497, + 0.014126078225672245, + 0.011676928959786892, + 0.015334222465753555, + 0.14916670322418213, + 0.0979514867067337, + 0.020062923431396484, + 0.008376166224479675, + 0.0025384409818798304, + 0.0034023160114884377, + 0.002010069089010358, + 0.0027654930017888546, + 0.004148596897721291, + 0.004492711275815964, + 0.008448783308267593, + 0.006359902676194906, + 0.0027819485403597355, + 0.002877345774322748, + 0.003283189609646797, + 0.0029322276823222637, + 0.01136750914156437, + 0.00994446687400341, + 0.00892488844692707, + 0.0044646840542554855, + 0.003953204955905676, + 0.004432107321918011, + 0.01856095716357231, + 0.004309020936489105, + 0.021843675523996353, + 0.008429187349975109, + 0.0033083728048950434, + 0.002628365997225046, + 0.007311458699405193, + 0.006285848096013069, + 0.008359004743397236, + 0.007205612491816282, + 0.016764214262366295, + 0.010956832207739353, + 0.012417508289217949, + 0.020712027326226234, + 0.0024473010562360287, + 0.002365631051361561, + 0.0022605792619287968, + 0.0028414269909262657, + 0.002973844762891531, + 0.003792412346228957, + 0.002056364668533206, + 0.0022564013488590717, + 0.00679708132520318, + 0.0073826489970088005, + 0.0043510254472494125, + 0.00415209261700511, + 0.004407577216625214, + 0.004263786599040031, + 0.002936446340754628, + 0.0037069383542984724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0028329261112958193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256996490061283 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005256996490061283, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256996490061283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794690489768982, + 0.11669205874204636, + 0.050776802003383636, + 0.17356108129024506, + 0.11518464237451553, + 0.05373026058077812, + 0.05526873841881752, + 0.14704614877700806, + 0.04360940307378769, + 0.06466186046600342 + ], + "credit_map": [ + 0.020161094143986702, + 0.022826489061117172, + 0.03522980958223343, + 0.019766544923186302, + 0.03139878436923027, + 0.019574128091335297, + 0.014490324072539806, + 0.01602187193930149, + 0.01691266894340515, + 0.02321917749941349, + 0.011883345432579517, + 0.01088770106434822, + 0.014724443666636944, + 0.01432829163968563, + 0.009545338340103626, + 0.015191095881164074, + 0.005995351821184158, + 0.005201796069741249, + 0.00367101957090199, + 0.005196577403694391, + 0.007016296964138746, + 0.012470564804971218, + 0.003644385142251849, + 0.007580810226500034, + 0.0081047173589468, + 0.008831929415464401, + 0.006652351468801498, + 0.0075847068801522255, + 0.07814626395702362, + 0.03171923756599426, + 0.02662898786365986, + 0.005892875604331493, + 0.009961814619600773, + 0.013055114075541496, + 0.012547415681183338, + 0.010272631421685219, + 0.012385090813040733, + 0.018140412867069244, + 0.02205258421599865, + 0.01676957495510578, + 0.002549988916143775, + 0.0027327865827828646, + 0.008382264524698257, + 0.00533206295222044, + 0.014954893849790096, + 0.009374392218887806, + 0.007489933632314205, + 0.002913939068093896, + 0.0039741359651088715, + 0.0040132333524525166, + 0.013370911590754986, + 0.004720790311694145, + 0.014572869054973125, + 0.0073245312087237835, + 0.0037950477562844753, + 0.0034972222056239843, + 0.015024839900434017, + 0.008835953660309315, + 0.011385845951735973, + 0.009684409946203232, + 0.027743449434638023, + 0.01786387898027897, + 0.01761295087635517, + 0.03889482095837593, + 0.005844887346029282, + 0.005241785664111376, + 0.003862819168716669, + 0.0054104323498904705, + 0.007263186387717724, + 0.007998147048056126, + 0.0034433549735695124, + 0.004544789437204599, + 0.011038859374821186, + 0.008987710811197758, + 0.0058914776891469955, + 0.007547713350504637, + 0.007969897240400314, + 0.01164665911346674, + 0.004596525337547064, + 0.006983019411563873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005256996490061283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774359479546547, + 0.05739918351173401, + 0.03985338658094406, + 0.4269583821296692, + 0.11722051352262497, + 0.08852653950452805, + 0.062024377286434174, + 0.07612060010433197, + 0.02958855964243412, + 0.0248725488781929 + ], + "credit_map": [ + 0.01115600485354662, + 0.007564415689557791, + 0.012275826185941696, + 0.009716428816318512, + 0.011606695130467415, + 0.00923824217170477, + 0.007232236210256815, + 0.008646093308925629, + 0.005870957858860493, + 0.006950530223548412, + 0.00805108156055212, + 0.005995688959956169, + 0.008634649217128754, + 0.007019294425845146, + 0.005343622528016567, + 0.009533359669148922, + 0.004152857232838869, + 0.003981047309935093, + 0.00812770240008831, + 0.004230892285704613, + 0.004605273250490427, + 0.007299453020095825, + 0.0028370514046400785, + 0.004619110841304064, + 0.01844632439315319, + 0.021363405510783195, + 0.014337560161948204, + 0.019237838685512543, + 0.19409886002540588, + 0.07604922354221344, + 0.06940987706184387, + 0.014015293680131435, + 0.010240592062473297, + 0.012343148700892925, + 0.01708565466105938, + 0.009968706406652927, + 0.022450802847743034, + 0.020248491317033768, + 0.011828288435935974, + 0.013054830953478813, + 0.005915218964219093, + 0.004933745600283146, + 0.015845036134123802, + 0.008169703185558319, + 0.029280371963977814, + 0.011181201785802841, + 0.006415183190256357, + 0.006786076817661524, + 0.0034247187431901693, + 0.004442353267222643, + 0.016409654170274734, + 0.004908392671495676, + 0.018958866596221924, + 0.007532484829425812, + 0.0034135763999074697, + 0.002934329677373171, + 0.006043501663953066, + 0.004693980794399977, + 0.007440787274390459, + 0.005407464224845171, + 0.013028571382164955, + 0.009521943517029285, + 0.010441414080560207, + 0.019542932510375977, + 0.0036929401103407145, + 0.004208363592624664, + 0.0030958554707467556, + 0.003778263693675399, + 0.0031696176156401634, + 0.004512273706495762, + 0.0026489936280995607, + 0.004482250194996595, + 0.0040572830475866795, + 0.0034995998721569777, + 0.0017169773345813155, + 0.0024702211376279593, + 0.0029705434571951628, + 0.005535470321774483, + 0.0018097584834322333, + 0.002812695689499378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.012535706162452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, smooth drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are giant Amazonian water lily leaves visible floating on a crystal-clear lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display curled upward edges, radiating veins, and a waxy sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and water surface exhibit a gentle, rhythmic rise and fall motion caused by a breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by emerald green and deep lake blue colors with shimmering golden highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015128131490200758 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.0015128131490200758, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015128131490200758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17349782586097717, + 0.06096817925572395, + 0.11371392756700516, + 0.19593599438667297, + 0.1633685827255249, + 0.04940374940633774, + 0.06696387380361557, + 0.09300021827220917, + 0.039231542497873306, + 0.04391610622406006 + ], + "credit_map": [ + 0.019851328805088997, + 0.021858103573322296, + 0.017724821344017982, + 0.02392992004752159, + 0.03165394812822342, + 0.02160179801285267, + 0.01754111424088478, + 0.019336795434355736, + 0.008746491745114326, + 0.009462276473641396, + 0.006139921955764294, + 0.007567528169602156, + 0.008255915716290474, + 0.007553268689662218, + 0.005542634986341, + 0.007700138725340366, + 0.00883946381509304, + 0.01039782166481018, + 0.01142210140824318, + 0.014168784953653812, + 0.014492455869913101, + 0.030119262635707855, + 0.008502038195729256, + 0.015772001817822456, + 0.012995642609894276, + 0.01134115643799305, + 0.009010719135403633, + 0.010130180045962334, + 0.073715440928936, + 0.030808471143245697, + 0.03890714794397354, + 0.00902724452316761, + 0.010640396736562252, + 0.01482503954321146, + 0.015117528848350048, + 0.01271999254822731, + 0.030957285314798355, + 0.02112104743719101, + 0.037781767547130585, + 0.020205534994602203, + 0.003238680772483349, + 0.002551307436078787, + 0.012736876495182514, + 0.0038436194881796837, + 0.012264675460755825, + 0.007362075615674257, + 0.0049165114760398865, + 0.0024900028947740793, + 0.007943828590214252, + 0.004165137652307749, + 0.014820259995758533, + 0.004237207118421793, + 0.02100253477692604, + 0.007833819836378098, + 0.003430276410654187, + 0.003530810121446848, + 0.012069725431501865, + 0.006309137679636478, + 0.004937788937240839, + 0.0071632592007517815, + 0.020867949351668358, + 0.011636782437562943, + 0.00954614207148552, + 0.020469430834054947, + 0.006801125127822161, + 0.004292965400964022, + 0.0026644638273864985, + 0.0036645811051130295, + 0.006034573074430227, + 0.008552570827305317, + 0.0033438564278185368, + 0.003877406707033515, + 0.008967462927103043, + 0.007005323190242052, + 0.004495424218475819, + 0.003527246182784438, + 0.0067116739228367805, + 0.006997387856245041, + 0.0029632593505084515, + 0.0032483309041708708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0015128131490200758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_tensor([8743, 8743])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_tensor([8743, 8743])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0623cc1651129d28a109075c2f62e4e639f0ee7d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_tensor([8743, 8743])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, majestic crane shot begins at a low angle, sweeping gracefully upward to reveal the breathtaking scale of a massive, multi-story greenhouse, rendered in a surreal and detailed anime-style composition. The structure, a soaring glass atrium and intricate white steel grid, creates a botanical cathedral. Inside, the air is filled with a soft, diffuse light. Vertical towers of luminous hydroponic plants rise from the polished concrete floor, their sleek frames supporting layers of lush, vibrant foliage, emitting a soft, ethereal teal glow. This cool interior glow contrasts beautifully with the warm golden rays of early morning sunlight filtering through the glass dome, casting long, dramatic rays of divine light that pierce the humid atmosphere. In classic Makoto Shinkai aesthetic, the entire scene is imbued with a sense of serene wonder, with countless glistening droplets of morning dew clinging to every leaf and glass with perfect surface tension. As the camera rises, these droplets catch the intersecting light sources, erupting in bright, sparkling anamorphic lens flares that dance across the frame, imbuing the futuristic yet natural environment with a palpable, shimmering energy.", + "seed": 15788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15788_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0447238449472934, + "vlm_logprob": -0.0447238449472934, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.0447238449472934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021088773384690285 + ], + "opposite_logprob_tokens": [ + -3.8960888385772705 + ], + "logprob_reward": -0.021088773384690285, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021088773384690285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07090911269187927, + 0.04117244854569435, + 0.025781741365790367, + 0.04851481691002846, + 0.022169316187500954, + 0.05245361104607582, + 0.353198766708374, + 0.206570565700531, + 0.1350771188735962, + 0.044152431190013885 + ], + "credit_map": [ + 0.006080230697989464, + 0.010375291109085083, + 0.009937967173755169, + 0.011650227941572666, + 0.006486775819212198, + 0.0063322787173092365, + 0.010787748731672764, + 0.009258594363927841, + 0.003686330048367381, + 0.008403240703046322, + 0.00568576343357563, + 0.004243663977831602, + 0.003261426230892539, + 0.003487922251224518, + 0.005674825049936771, + 0.006729277316480875, + 0.0031923940405249596, + 0.002298501553013921, + 0.002818439854308963, + 0.0029142487328499556, + 0.0029355839360505342, + 0.004192816559225321, + 0.004211898427456617, + 0.003217859659343958, + 0.007520620245486498, + 0.006442129611968994, + 0.004743748810142279, + 0.004583948291838169, + 0.006213818211108446, + 0.008522704243659973, + 0.00571725657209754, + 0.004770589526742697, + 0.0035968495067209005, + 0.0035526291467249393, + 0.0038540977984666824, + 0.00212503201328218, + 0.002291515003889799, + 0.0026019709184765816, + 0.002168891252949834, + 0.001978330546990037, + 0.005404544062912464, + 0.0065596685744822025, + 0.004889290314167738, + 0.009461832232773304, + 0.0055904146283864975, + 0.006577099673449993, + 0.008701973594725132, + 0.0052687861025333405, + 0.11731627583503723, + 0.1416737586259842, + 0.011548198759555817, + 0.008254878222942352, + 0.034362196922302246, + 0.02099987305700779, + 0.009276889264583588, + 0.009766694158315659, + 0.031406156718730927, + 0.03218461200594902, + 0.016895035281777382, + 0.012270963750779629, + 0.03110789693892002, + 0.04645144194364548, + 0.01321372203528881, + 0.023040741682052612, + 0.014856764115393162, + 0.017721164971590042, + 0.012537844479084015, + 0.014012040570378304, + 0.0194869302213192, + 0.024207253009080887, + 0.018136033788323402, + 0.014119076542556286, + 0.00494808005169034, + 0.005208848509937525, + 0.004473394714295864, + 0.0050341542810201645, + 0.007946795783936977, + 0.006091955583542585, + 0.005037013441324234, + 0.005412189289927483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.021088773384690285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036048531532288 + ], + "opposite_logprob_tokens": [ + -1.9103604555130005 + ], + "logprob_reward": -0.16036048531532288, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036048531532288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07080118358135223, + 0.025787288323044777, + 0.030834266915917397, + 0.05139188468456268, + 0.070060595870018, + 0.042034655809402466, + 0.04438862204551697, + 0.36917591094970703, + 0.19318589568138123, + 0.10233970731496811 + ], + "credit_map": [ + 0.00846505630761385, + 0.007763307075947523, + 0.01141496654599905, + 0.008704347535967827, + 0.0063919974491000175, + 0.008562635630369186, + 0.007718150038272142, + 0.011780722066760063, + 0.002659895922988653, + 0.005670828744769096, + 0.00277898833155632, + 0.0025591086596250534, + 0.0027043104637414217, + 0.003341963514685631, + 0.003062801668420434, + 0.00300939055159688, + 0.0035508633591234684, + 0.0028077391907572746, + 0.0034618324134498835, + 0.005741455592215061, + 0.002776769222691655, + 0.003807314205914736, + 0.005250787362456322, + 0.003437506267800927, + 0.008947433903813362, + 0.008714323863387108, + 0.005571714602410793, + 0.005090611055493355, + 0.004467211663722992, + 0.006646921392530203, + 0.005591605324298143, + 0.006362064275890589, + 0.011187296360731125, + 0.0125364875420928, + 0.012302200309932232, + 0.006981902755796909, + 0.006139214616268873, + 0.008391402661800385, + 0.006565837189555168, + 0.005956255365163088, + 0.006046953145414591, + 0.008043159730732441, + 0.003814131487160921, + 0.0034766346216201782, + 0.0036484082229435444, + 0.00537376943975687, + 0.006357260979712009, + 0.005274342373013496, + 0.010315309278666973, + 0.011124095879495144, + 0.004060585517436266, + 0.002654843730852008, + 0.004272168967872858, + 0.004822538234293461, + 0.003919853363186121, + 0.003219227073714137, + 0.05128118023276329, + 0.05610908195376396, + 0.03272949531674385, + 0.025954751297831535, + 0.05166177079081535, + 0.08160312473773956, + 0.025375444442033768, + 0.04446105286478996, + 0.020474057644605637, + 0.026909179985523224, + 0.013901449739933014, + 0.021063746884465218, + 0.02912375144660473, + 0.030692754313349724, + 0.03286257013678551, + 0.01815837435424328, + 0.012070602737367153, + 0.012056629173457623, + 0.011249803937971592, + 0.011640451848506927, + 0.01698336750268936, + 0.015499608591198921, + 0.010550483129918575, + 0.012288765050470829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.16036048531532288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955214146524668 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955214146524668, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955214146524668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14338113367557526, + 0.1223655492067337, + 0.14776161313056946, + 0.10876046866178513, + 0.05325673893094063, + 0.05524853989481926, + 0.09786492586135864, + 0.10043489187955856, + 0.07992462068796158, + 0.09100155532360077 + ], + "credit_map": [ + 0.01726393587887287, + 0.016999445855617523, + 0.024196242913603783, + 0.015440915711224079, + 0.013801776804029942, + 0.01689102128148079, + 0.017456982284784317, + 0.021330805495381355, + 0.00903843529522419, + 0.010603433474898338, + 0.02331809140741825, + 0.01537324395030737, + 0.009928428567945957, + 0.009426995180547237, + 0.016184309497475624, + 0.028492605313658714, + 0.02094436064362526, + 0.01547672227025032, + 0.01658022776246071, + 0.02056342549622059, + 0.011776627041399479, + 0.026457395404577255, + 0.017283257097005844, + 0.01867958903312683, + 0.013226835057139397, + 0.014871407300233841, + 0.011909534223377705, + 0.009249159134924412, + 0.011890340596437454, + 0.016248302534222603, + 0.01781163364648819, + 0.013553248718380928, + 0.010085845366120338, + 0.011682381853461266, + 0.005529263988137245, + 0.005395365413278341, + 0.0054730381816625595, + 0.0068518673069775105, + 0.004288500640541315, + 0.003950473852455616, + 0.008706094697117805, + 0.01404685527086258, + 0.0038078948855400085, + 0.0044705308973789215, + 0.006966202519834042, + 0.007157079875469208, + 0.005657694302499294, + 0.0044361865147948265, + 0.028209956362843513, + 0.03293467313051224, + 0.005166690796613693, + 0.0038243692833930254, + 0.008532511070370674, + 0.008033453486859798, + 0.00732625974342227, + 0.003837019205093384, + 0.013463569805026054, + 0.013713990338146687, + 0.009642882272601128, + 0.009258043952286243, + 0.019546668976545334, + 0.013143622316420078, + 0.008916131220757961, + 0.01274997927248478, + 0.007864723913371563, + 0.008956129662692547, + 0.00967799499630928, + 0.012596157379448414, + 0.011513280682265759, + 0.009641262702643871, + 0.012911958619952202, + 0.006763111334294081, + 0.010527906939387321, + 0.013059807009994984, + 0.007382742594927549, + 0.009092620573937893, + 0.017336681485176086, + 0.011361144483089447, + 0.010206619277596474, + 0.0120340371504426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005955214146524668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372518718242645 + ], + "opposite_logprob_tokens": [ + -3.4087252616882324 + ], + "logprob_reward": -0.03372518718242645, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372518718242645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08056563884019852, + 0.03701784834265709, + 0.029468439519405365, + 0.046177711337804794, + 0.11573457717895508, + 0.16599810123443604, + 0.3216022551059723, + 0.08964404463768005, + 0.044822365045547485, + 0.0689689889550209 + ], + "credit_map": [ + 0.008478071540594101, + 0.011769167147576809, + 0.008720568381249905, + 0.011183018796145916, + 0.008191194385290146, + 0.00980981346219778, + 0.012788455002009869, + 0.009625354781746864, + 0.004476409405469894, + 0.007029310800135136, + 0.005174476653337479, + 0.004006878938525915, + 0.002808242104947567, + 0.0036688547115772963, + 0.005614512600004673, + 0.0042391614988446236, + 0.003890742314979434, + 0.0032525162678211927, + 0.00355161027982831, + 0.002956621116027236, + 0.0035728972870856524, + 0.00530235143378377, + 0.003900267882272601, + 0.0030414338689297438, + 0.006451509892940521, + 0.0076477788388729095, + 0.00443964172154665, + 0.0036843244452029467, + 0.005681987851858139, + 0.0059749772772192955, + 0.005644990596920252, + 0.006652499083429575, + 0.01858467608690262, + 0.01717321202158928, + 0.018361054360866547, + 0.011323806829750538, + 0.0121559863910079, + 0.01571706123650074, + 0.011726046912372112, + 0.010692735202610493, + 0.018414577469229698, + 0.021282590925693512, + 0.01627998799085617, + 0.031149758026003838, + 0.016141260042786598, + 0.017604688182473183, + 0.02928883768618107, + 0.015836413949728012, + 0.11401841789484024, + 0.12398749589920044, + 0.010186835192143917, + 0.0075263334438204765, + 0.03214232623577118, + 0.014563709497451782, + 0.010563530027866364, + 0.008613592945039272, + 0.014570103026926517, + 0.014563611708581448, + 0.007025919388979673, + 0.0066606756299734116, + 0.017883647233247757, + 0.011860262602567673, + 0.00931637268513441, + 0.007763453293591738, + 0.005338629707694054, + 0.005487618036568165, + 0.005057218950241804, + 0.0069917780347168446, + 0.007192528806626797, + 0.006073267664760351, + 0.004721041768789291, + 0.003960277885198593, + 0.006457549519836903, + 0.00845930352807045, + 0.005297387484461069, + 0.012171097099781036, + 0.010091163218021393, + 0.006466648541390896, + 0.009555242024362087, + 0.010470602661371231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.03372518718242645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12164003401994705, + 0.06436452269554138, + 0.02895275689661503, + 0.039973169565200806, + 0.05041281133890152, + 0.06326590478420258, + 0.12338747084140778, + 0.23476599156856537, + 0.18053023517131805, + 0.09270717948675156 + ], + "credit_map": [ + 0.014894973486661911, + 0.01895149238407612, + 0.01294547040015459, + 0.016193363815546036, + 0.014660196378827095, + 0.015382243320345879, + 0.017187289893627167, + 0.011424989439547062, + 0.0054058716632425785, + 0.010179135017096996, + 0.010696484707295895, + 0.005441876593977213, + 0.004040275700390339, + 0.005095343105494976, + 0.010137651115655899, + 0.013367877341806889, + 0.0041353702545166016, + 0.003207483794540167, + 0.0035983333364129066, + 0.0036613706033676863, + 0.0036306262481957674, + 0.0035352702252566814, + 0.004161774180829525, + 0.0030225247610360384, + 0.007334205321967602, + 0.007128874305635691, + 0.0031679801177233458, + 0.0026806476525962353, + 0.005305117461830378, + 0.005896962713450193, + 0.0044924248941242695, + 0.003966956865042448, + 0.009631151333451271, + 0.007426298223435879, + 0.008410446345806122, + 0.0046838633716106415, + 0.005117253866046667, + 0.006460445933043957, + 0.004528947174549103, + 0.004154404625296593, + 0.009152118116617203, + 0.014700003899633884, + 0.00547810597345233, + 0.007127129007130861, + 0.007745112758129835, + 0.007666900288313627, + 0.006849454250186682, + 0.004547078162431717, + 0.0372224822640419, + 0.035066284239292145, + 0.007494899444282055, + 0.005097329616546631, + 0.014143065549433231, + 0.010258786380290985, + 0.009744826704263687, + 0.0043597836047410965, + 0.03066669963300228, + 0.03440062701702118, + 0.019874751567840576, + 0.016323965042829514, + 0.03590349853038788, + 0.05303172394633293, + 0.019232649356126785, + 0.02533205971121788, + 0.01976015605032444, + 0.023338032886385918, + 0.013935891911387444, + 0.014721729792654514, + 0.029917074367403984, + 0.020393377169966698, + 0.04295574501156807, + 0.015508206561207771, + 0.011740626767277718, + 0.014148177579045296, + 0.008103613741695881, + 0.009073818102478981, + 0.021582888439297676, + 0.008607507683336735, + 0.01042377483099699, + 0.009026758372783661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0024895647075027227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic crane shot begins at a low angle, sweeping gracefully upward to reveal the breathtaking scale of a massive, multi-story greenhouse, rendered in a surreal and detailed anime-style composition. The structure, a soaring glass atrium and intricate white steel grid, creates a botanical cathedral. Inside, the air is filled with a soft, diffuse light. Vertical towers of luminous hydroponic plants rise from the polished concrete floor, their sleek frames supporting layers of lush, vibrant foliage, emitting a soft, ethereal teal glow. This cool interior glow contrasts beautifully with the warm golden rays of early morning sunlight filtering through the glass dome, casting long, dramatic rays of divine light that pierce the humid atmosphere. In classic Makoto Shinkai aesthetic, the entire scene is imbued with a sense of serene wonder, with countless glistening droplets of morning dew clinging to every leaf and glass with perfect surface tension. As the camera rises, these droplets catch the intersecting light sources, erupting in bright, sparkling anamorphic lens flares that dance across the frame, imbuing the futuristic yet natural environment with a palpable, shimmering energy.", + "seed": 15789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15789_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11737280343659222, + "vlm_logprob": -0.11737280343659222, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.11737280343659222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1280970573425293 + ], + "opposite_logprob_tokens": [ + -2.1280970573425293 + ], + "logprob_reward": -0.1280970573425293, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1280970573425293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22553712129592896, + 0.028362765908241272, + 0.06362376362085342, + 0.05454018712043762, + 0.034358490258455276, + 0.04898142069578171, + 0.28162676095962524, + 0.03803739324212074, + 0.08174023032188416, + 0.143191859126091 + ], + "credit_map": [ + 0.008611892350018024, + 0.011447281576693058, + 0.012653209269046783, + 0.07505366206169128, + 0.014360793866217136, + 0.010772118344902992, + 0.014815046451985836, + 0.07782311737537384, + 0.004064691253006458, + 0.0048296344466507435, + 0.002627368550747633, + 0.0025536841712892056, + 0.0034339474514126778, + 0.004867719952017069, + 0.0029843440279364586, + 0.003001378383487463, + 0.0067307972349226475, + 0.011592748574912548, + 0.005269428249448538, + 0.004132934380322695, + 0.012897085398435593, + 0.014218656346201897, + 0.004650036338716745, + 0.0041320836171507835, + 0.0040823593735694885, + 0.006096136290580034, + 0.003393008606508374, + 0.002984548220410943, + 0.014687992632389069, + 0.016114719212055206, + 0.003907682839781046, + 0.0032737439032644033, + 0.0034401975572109222, + 0.0044748480431735516, + 0.0027759436052292585, + 0.005074354354292154, + 0.004553842358291149, + 0.005066786427050829, + 0.006140395998954773, + 0.002832124475389719, + 0.0037904796190559864, + 0.006313755642622709, + 0.005741245578974485, + 0.0074034868739545345, + 0.0065162996761500835, + 0.006365851499140263, + 0.007093953434377909, + 0.0057563516311347485, + 0.034818198531866074, + 0.04437275230884552, + 0.017954306676983833, + 0.018201783299446106, + 0.05802471190690994, + 0.05747435241937637, + 0.02525777742266655, + 0.025522874668240547, + 0.004797739442437887, + 0.00508915213868022, + 0.0032604059670120478, + 0.005087824072688818, + 0.008433718234300613, + 0.00442481366917491, + 0.0035792896524071693, + 0.0033644530922174454, + 0.007230442948639393, + 0.009173198603093624, + 0.012719951570034027, + 0.0118597112596035, + 0.012746191583573818, + 0.012172841466963291, + 0.00754900649189949, + 0.00828888826072216, + 0.015271061100065708, + 0.022361166775226593, + 0.010919841006398201, + 0.02502794750034809, + 0.025717200711369514, + 0.013361750170588493, + 0.014566156081855297, + 0.01596674509346485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.1280970573425293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42886555194854736 + ], + "opposite_logprob_tokens": [ + -1.0538655519485474 + ], + "logprob_reward": -0.42886555194854736, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42886555194854736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23998966813087463, + 0.08205198496580124, + 0.07428494840860367, + 0.03822791948914528, + 0.02301741950213909, + 0.11831871420145035, + 0.1359466314315796, + 0.05668782442808151, + 0.045513831079006195, + 0.1859610676765442 + ], + "credit_map": [ + 0.011981235817074776, + 0.016053928062319756, + 0.01880704052746296, + 0.06468140333890915, + 0.030245481058955193, + 0.0184203889220953, + 0.021812908351421356, + 0.0579872690141201, + 0.005562573671340942, + 0.007274136878550053, + 0.00815140176564455, + 0.016015799716114998, + 0.008407354354858398, + 0.008350411430001259, + 0.01937755011022091, + 0.008912750519812107, + 0.006872309371829033, + 0.017686057835817337, + 0.00563106220215559, + 0.006230476312339306, + 0.006959804333746433, + 0.017250632867217064, + 0.007224540691822767, + 0.006430056877434254, + 0.003330116393044591, + 0.006376111414283514, + 0.00288609997369349, + 0.0038488421123474836, + 0.0068561541847884655, + 0.007344126235693693, + 0.004193315282464027, + 0.0033931536599993706, + 0.0021274976897984743, + 0.0032948730513453484, + 0.0023045476991683245, + 0.004144076723605394, + 0.0028638869989663363, + 0.0024871535133570433, + 0.002971224021166563, + 0.002824159571900964, + 0.011528637260198593, + 0.015870612114667892, + 0.014371906407177448, + 0.0171964131295681, + 0.014535820111632347, + 0.012156658805906773, + 0.014284882694482803, + 0.018373781815171242, + 0.0138015728443861, + 0.02104700729250908, + 0.009763801470398903, + 0.01355601754039526, + 0.0263010635972023, + 0.029973836615681648, + 0.010413775220513344, + 0.011089545674622059, + 0.005904104094952345, + 0.007304848171770573, + 0.007098667323589325, + 0.007749198004603386, + 0.00820634514093399, + 0.00908629409968853, + 0.004366687033325434, + 0.006971675902605057, + 0.0048770117573440075, + 0.00527977617457509, + 0.006789028644561768, + 0.006132848095148802, + 0.005445622373372316, + 0.004614728968590498, + 0.005143394693732262, + 0.007231418509036303, + 0.016501260921359062, + 0.0271401759237051, + 0.012752518057823181, + 0.0586521215736866, + 0.022152163088321686, + 0.011562817730009556, + 0.022585883736610413, + 0.014614105224609375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.42886555194854736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410179328173399 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.00410179328173399, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410179328173399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23642070591449738, + 0.23786717653274536, + 0.05960964784026146, + 0.07002656161785126, + 0.02195044606924057, + 0.09545499086380005, + 0.10806036740541458, + 0.04313083365559578, + 0.0415886789560318, + 0.08589056134223938 + ], + "credit_map": [ + 0.013374599628150463, + 0.015638450160622597, + 0.02013344131410122, + 0.05492551624774933, + 0.02524072304368019, + 0.019854888319969177, + 0.02551046945154667, + 0.061742622405290604, + 0.026504836976528168, + 0.02213476225733757, + 0.0269624050706625, + 0.035912491381168365, + 0.02181774377822876, + 0.02920139580965042, + 0.04465571045875549, + 0.03067781776189804, + 0.007358888629823923, + 0.011871074326336384, + 0.004432625602930784, + 0.006601276807487011, + 0.007452546618878841, + 0.013249470852315426, + 0.004091632552444935, + 0.004552136175334454, + 0.00458397064357996, + 0.011965423822402954, + 0.0038165184669196606, + 0.005243919789791107, + 0.01860499195754528, + 0.017212973907589912, + 0.004895336460322142, + 0.0037034351844340563, + 0.0019641011022031307, + 0.0034749251790344715, + 0.002048988128080964, + 0.003770449198782444, + 0.002574506914243102, + 0.0031554843299090862, + 0.002793404273688793, + 0.002168586477637291, + 0.0094728022813797, + 0.011395607143640518, + 0.013353511691093445, + 0.013678252696990967, + 0.010478684678673744, + 0.009538261219859123, + 0.011018442921340466, + 0.01651942729949951, + 0.010107530280947685, + 0.013841836713254452, + 0.010061826556921005, + 0.016092296689748764, + 0.016874821856617928, + 0.020968226715922356, + 0.01069046463817358, + 0.009423371404409409, + 0.004555673338472843, + 0.0055305385030806065, + 0.004427907522767782, + 0.008399520069360733, + 0.0059820981696248055, + 0.00530096422880888, + 0.004263032227754593, + 0.004671098664402962, + 0.005007131490856409, + 0.0048775095492601395, + 0.005745995789766312, + 0.006105801090598106, + 0.005526384338736534, + 0.005799977108836174, + 0.00384667981415987, + 0.0046791997738182545, + 0.006171060260385275, + 0.011478472501039505, + 0.0071540093049407005, + 0.025244763121008873, + 0.010671666823327541, + 0.006103680934756994, + 0.010590700432658195, + 0.00847620889544487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.00410179328173399 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330612577497959 + ], + "opposite_logprob_tokens": [ + -3.773306131362915 + ], + "logprob_reward": -0.02330612577497959, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330612577497959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2270417958498001, + 0.09562741965055466, + 0.04556858167052269, + 0.03929808735847473, + 0.056933678686618805, + 0.14117440581321716, + 0.10757090896368027, + 0.11562483757734299, + 0.0671958476305008, + 0.1039644107222557 + ], + "credit_map": [ + 0.01030066516250372, + 0.013298538513481617, + 0.01996680162847042, + 0.06804132461547852, + 0.01584560237824917, + 0.013934966176748276, + 0.021999675780534744, + 0.06365423649549484, + 0.01028541661798954, + 0.0103971092030406, + 0.012407323345541954, + 0.009227059781551361, + 0.013663639314472675, + 0.014243091456592083, + 0.011000057682394981, + 0.014403731562197208, + 0.004857942461967468, + 0.00578699167817831, + 0.006585629191249609, + 0.007546707522124052, + 0.006350584328174591, + 0.006983826402574778, + 0.004395599011331797, + 0.003061300376430154, + 0.003362581366673112, + 0.0037658039946109056, + 0.002737519098445773, + 0.0029627715703099966, + 0.009996721521019936, + 0.010950994677841663, + 0.002624763874337077, + 0.0028969331178814173, + 0.00798968505114317, + 0.008025191724300385, + 0.0041294037364423275, + 0.0070721060037612915, + 0.011436973698437214, + 0.006435472052544355, + 0.00719560869038105, + 0.004649240989238024, + 0.013990555889904499, + 0.01593242771923542, + 0.023184344172477722, + 0.022892480716109276, + 0.016139740124344826, + 0.013855662196874619, + 0.01625613123178482, + 0.018923062831163406, + 0.010272476822137833, + 0.015768475830554962, + 0.0073738484643399715, + 0.011578349396586418, + 0.019935712218284607, + 0.023864632472395897, + 0.008947375230491161, + 0.009830036200582981, + 0.009577234275639057, + 0.014006705954670906, + 0.010228349827229977, + 0.032339908182621, + 0.017016546800732613, + 0.008942637592554092, + 0.013558325357735157, + 0.009955131448805332, + 0.005638403352349997, + 0.008480933494865894, + 0.011058705858886242, + 0.010360955260694027, + 0.011586030945181847, + 0.0077316961251199245, + 0.005962023511528969, + 0.006377103738486767, + 0.008026541210711002, + 0.01442949939519167, + 0.011276695877313614, + 0.023105675354599953, + 0.0148998461663723, + 0.006867911666631699, + 0.015927234664559364, + 0.00943100918084383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.02330612577497959 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493488835170865 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493488835170865, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493488835170865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19232212007045746, + 0.06533531844615936, + 0.044213417917490005, + 0.05115429684519768, + 0.0640535056591034, + 0.11287906765937805, + 0.10483244061470032, + 0.051825959235429764, + 0.08362975716590881, + 0.22975414991378784 + ], + "credit_map": [ + 0.01283883024007082, + 0.012258379720151424, + 0.019722377881407738, + 0.03985558822751045, + 0.014152317307889462, + 0.018024617806077003, + 0.024216096848249435, + 0.051253899931907654, + 0.006611529737710953, + 0.00752673577517271, + 0.005619319621473551, + 0.00787793193012476, + 0.013234754092991352, + 0.010967624373733997, + 0.005637568887323141, + 0.007859847508370876, + 0.005248707253485918, + 0.0061934576369822025, + 0.006020034197717905, + 0.0066143530420959, + 0.00490359915420413, + 0.006338007282465696, + 0.004193675238639116, + 0.00470158364623785, + 0.00415410753339529, + 0.009171983227133751, + 0.002630223985761404, + 0.0034067293163388968, + 0.012505915947258472, + 0.012083332054316998, + 0.0038795906584709883, + 0.003322406206279993, + 0.003919254057109356, + 0.006892789620906115, + 0.00498129753395915, + 0.016017602756619453, + 0.0066717215813696384, + 0.005623192526400089, + 0.013954139314591885, + 0.005993505008518696, + 0.009353401139378548, + 0.017234114930033684, + 0.012173047289252281, + 0.01724548637866974, + 0.016109323129057884, + 0.009969128295779228, + 0.013720187358558178, + 0.01707436703145504, + 0.011825064197182655, + 0.015629470348358154, + 0.00832322146743536, + 0.010064059868454933, + 0.0204012431204319, + 0.021661950275301933, + 0.008079444989562035, + 0.008847990073263645, + 0.004567044787108898, + 0.0075242649763822556, + 0.005011474713683128, + 0.005403558723628521, + 0.010959617793560028, + 0.007206434849649668, + 0.004028795287013054, + 0.007124762982130051, + 0.0073060705326497555, + 0.01120778452605009, + 0.008892255835235119, + 0.011423517018556595, + 0.013541524298489094, + 0.006643414963036776, + 0.01497576292604208, + 0.009639414958655834, + 0.018833843991160393, + 0.03436622768640518, + 0.016772976145148277, + 0.07629704475402832, + 0.02390272542834282, + 0.014993532560765743, + 0.028426507487893105, + 0.016161276027560234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.002493488835170865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic crane shot begins at a low angle, sweeping gracefully upward to reveal the breathtaking scale of a massive, multi-story greenhouse, rendered in a surreal and detailed anime-style composition. The structure, a soaring glass atrium and intricate white steel grid, creates a botanical cathedral. Inside, the air is filled with a soft, diffuse light. Vertical towers of luminous hydroponic plants rise from the polished concrete floor, their sleek frames supporting layers of lush, vibrant foliage, emitting a soft, ethereal teal glow. This cool interior glow contrasts beautifully with the warm golden rays of early morning sunlight filtering through the glass dome, casting long, dramatic rays of divine light that pierce the humid atmosphere. In classic Makoto Shinkai aesthetic, the entire scene is imbued with a sense of serene wonder, with countless glistening droplets of morning dew clinging to every leaf and glass with perfect surface tension. As the camera rises, these droplets catch the intersecting light sources, erupting in bright, sparkling anamorphic lens flares that dance across the frame, imbuing the futuristic yet natural environment with a palpable, shimmering energy.", + "seed": 15790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15790_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028855423722416163, + "vlm_logprob": -0.028855423722416163, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.028855423722416163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0800245851278305 + ], + "opposite_logprob_tokens": [ + -2.580024480819702 + ], + "logprob_reward": -0.0800245851278305, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0800245851278305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08048778027296066, + 0.07755199074745178, + 0.05160866677761078, + 0.08572997152805328, + 0.02044733427464962, + 0.03927551209926605, + 0.09372659772634506, + 0.038263119757175446, + 0.08128625899553299, + 0.43162277340888977 + ], + "credit_map": [ + 0.006754316855221987, + 0.009158061817288399, + 0.007995769381523132, + 0.005449726711958647, + 0.007377387955784798, + 0.006692139431834221, + 0.01954510807991028, + 0.01751527190208435, + 0.006718784105032682, + 0.017525777220726013, + 0.0068832263350486755, + 0.006387671921402216, + 0.014016559347510338, + 0.010537912137806416, + 0.006880792789161205, + 0.00860126968473196, + 0.005913961213082075, + 0.005155487917363644, + 0.0037090873811393976, + 0.009652815759181976, + 0.006816243752837181, + 0.006588991731405258, + 0.006627053953707218, + 0.007145027630031109, + 0.018168576061725616, + 0.011205997318029404, + 0.0061632441356778145, + 0.01086832582950592, + 0.015389924868941307, + 0.0071261790581047535, + 0.007187688257545233, + 0.00962004717439413, + 0.0027520391158759594, + 0.0036407150328159332, + 0.002351092640310526, + 0.002144516445696354, + 0.003435904625803232, + 0.0019442641641944647, + 0.0016716785030439496, + 0.0025071243289858103, + 0.005762211978435516, + 0.004379292018711567, + 0.004688470624387264, + 0.00397614948451519, + 0.007648938801139593, + 0.0037092037964612246, + 0.004396646749228239, + 0.004714597947895527, + 0.013080093078315258, + 0.009277479723095894, + 0.0088130421936512, + 0.017726611346006393, + 0.015837082639336586, + 0.007571244612336159, + 0.010128656402230263, + 0.011292394250631332, + 0.004926351364701986, + 0.0037971760611981153, + 0.004236211534589529, + 0.004868735559284687, + 0.005765484180301428, + 0.0039691077545285225, + 0.005850465968251228, + 0.004849586170166731, + 0.006662552244961262, + 0.00936499610543251, + 0.011002043262124062, + 0.01287817768752575, + 0.012292571365833282, + 0.009173073805868626, + 0.012422834523022175, + 0.0074900067411363125, + 0.06499768793582916, + 0.06777989864349365, + 0.04543379694223404, + 0.04538995772600174, + 0.07404307276010513, + 0.05228438973426819, + 0.047727350145578384, + 0.03396664559841156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0800245851278305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073763370513916 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073763370513916, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073763370513916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3860631585121155, + 0.1071634590625763, + 0.09699289500713348, + 0.12337318062782288, + 0.03578021749854088, + 0.044359974563121796, + 0.024387933313846588, + 0.013199422508478165, + 0.036493923515081406, + 0.13218577206134796 + ], + "credit_map": [ + 0.032044991850852966, + 0.049371182918548584, + 0.038090988993644714, + 0.025127213448286057, + 0.07442037761211395, + 0.03793298453092575, + 0.05716145411133766, + 0.07191399484872818, + 0.008363853208720684, + 0.0257602259516716, + 0.00904734805226326, + 0.00713344058021903, + 0.025489548221230507, + 0.013865926302969456, + 0.007558284793049097, + 0.00994484219700098, + 0.010328452102839947, + 0.01321074552834034, + 0.006771148182451725, + 0.018926559016108513, + 0.012377329170703888, + 0.012424259446561337, + 0.011464901268482208, + 0.01148950308561325, + 0.0169217586517334, + 0.021636907011270523, + 0.010734432376921177, + 0.013352807611227036, + 0.010513339191675186, + 0.01881338655948639, + 0.013900178484618664, + 0.017500387504696846, + 0.004282514099031687, + 0.004136422649025917, + 0.003618129761889577, + 0.004069927614182234, + 0.006838718894869089, + 0.0042796870693564415, + 0.0032031559385359287, + 0.005351660307496786, + 0.006434084847569466, + 0.00636643823236227, + 0.0026991446502506733, + 0.003285530721768737, + 0.010757118463516235, + 0.004476778209209442, + 0.005003581289201975, + 0.005337301176041365, + 0.003306739265099168, + 0.0029694433324038982, + 0.0028382777236402035, + 0.004032624885439873, + 0.0043249367736279964, + 0.002406026003882289, + 0.0022637222427874804, + 0.0022461640182882547, + 0.001622545882128179, + 0.0014999710256233811, + 0.001529659959487617, + 0.0017608815105631948, + 0.0022972652222961187, + 0.0014628738863393664, + 0.0011100746924057603, + 0.0019161509117111564, + 0.0033779528457671404, + 0.004048412665724754, + 0.005894924513995647, + 0.005075729917734861, + 0.005771389231085777, + 0.004729082807898521, + 0.0043750121258199215, + 0.003221421968191862, + 0.013950170949101448, + 0.015864675864577293, + 0.01367270853370428, + 0.018144868314266205, + 0.02229829877614975, + 0.022884715348482132, + 0.014903281815350056, + 0.01046706736087799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.016073763370513916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6457111239433289, + 0.07150603830814362, + 0.04288652166724205, + 0.09621390700340271, + 0.02354353666305542, + 0.03622429445385933, + 0.014925997704267502, + 0.01934993825852871, + 0.027196699753403664, + 0.02244209125638008 + ], + "credit_map": [ + 0.036686066538095474, + 0.1605950891971588, + 0.03280412405729294, + 0.030292987823486328, + 0.07079962641000748, + 0.032733917236328125, + 0.18976867198944092, + 0.09203053265810013, + 0.0058253854513168335, + 0.016966335475444794, + 0.006841151975095272, + 0.004432493355125189, + 0.01494925282895565, + 0.010379143990576267, + 0.0048706624656915665, + 0.0072416020557284355, + 0.003928519319742918, + 0.004798384383320808, + 0.0024796733632683754, + 0.006812995299696922, + 0.006651103030890226, + 0.00846431776881218, + 0.004949565045535564, + 0.004801961127668619, + 0.009898524731397629, + 0.016676634550094604, + 0.006999536417424679, + 0.008863775990903378, + 0.011642455123364925, + 0.013808480463922024, + 0.010395500808954239, + 0.017928991466760635, + 0.003211516421288252, + 0.001995913917198777, + 0.002501067239791155, + 0.0019217249937355518, + 0.0030989230144768953, + 0.0031441154424101114, + 0.002578107640147209, + 0.005092165898531675, + 0.005755397956818342, + 0.006454378832131624, + 0.002496269764378667, + 0.0023454572074115276, + 0.01006929948925972, + 0.0025488180108368397, + 0.0038636927492916584, + 0.002690976718440652, + 0.0024874371010810137, + 0.001693430356681347, + 0.0013374187983572483, + 0.001983767142519355, + 0.0024563707411289215, + 0.0017655022675171494, + 0.0013404067140072584, + 0.0018616627203300595, + 0.0026884751860052347, + 0.0022206937428563833, + 0.003103084396570921, + 0.00237569073215127, + 0.002842491026967764, + 0.001999957486987114, + 0.002312561497092247, + 0.0018069833749905229, + 0.002488980069756508, + 0.0036951578222215176, + 0.0031179403886198997, + 0.003935209475457668, + 0.004638390149921179, + 0.004581149201840162, + 0.002923310501500964, + 0.0018165602814406157, + 0.0028203600086271763, + 0.0027352063916623592, + 0.0032648765482008457, + 0.0016969121061265469, + 0.003332478692755103, + 0.004363844636827707, + 0.002220351016148925, + 0.002008059062063694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812828287482262 + ], + "opposite_logprob_tokens": [ + -3.28812837600708 + ], + "logprob_reward": -0.03812828287482262, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812828287482262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2869530916213989, + 0.08967693150043488, + 0.09686306118965149, + 0.07317353039979935, + 0.05197393149137497, + 0.06799386441707611, + 0.04078378155827522, + 0.03901256248354912, + 0.06073809415102005, + 0.1928310990333557 + ], + "credit_map": [ + 0.03252418339252472, + 0.05574413761496544, + 0.047736525535583496, + 0.027118410915136337, + 0.04306063428521156, + 0.03478209674358368, + 0.023181775584816933, + 0.022805320098996162, + 0.008971590548753738, + 0.02031610533595085, + 0.008028904907405376, + 0.005642423871904612, + 0.020023606717586517, + 0.010246488265693188, + 0.007850871421396732, + 0.008596942760050297, + 0.009932764805853367, + 0.009980536065995693, + 0.005642687436193228, + 0.016376178711652756, + 0.015214809216558933, + 0.016695139929652214, + 0.011776896193623543, + 0.011244048364460468, + 0.007252249866724014, + 0.014352748170495033, + 0.004625428467988968, + 0.006661726627498865, + 0.01104073878377676, + 0.009323959238827229, + 0.007177991792559624, + 0.012738688848912716, + 0.006717556621879339, + 0.005066453013569117, + 0.005447732750326395, + 0.0046408954076468945, + 0.007434716913849115, + 0.007116714958101511, + 0.00792677327990532, + 0.007623090874403715, + 0.0123209897428751, + 0.007514643017202616, + 0.004646155517548323, + 0.007299920544028282, + 0.016297658905386925, + 0.004847649019211531, + 0.008091246709227562, + 0.006975600961595774, + 0.005647662095725536, + 0.004668057896196842, + 0.004013198893517256, + 0.0077746897004544735, + 0.006279362831264734, + 0.003721916116774082, + 0.004581792745739222, + 0.004097098484635353, + 0.005157687235623598, + 0.004984341096132994, + 0.0053184498101472855, + 0.0040666768327355385, + 0.007135884370654821, + 0.004846279975026846, + 0.004563506692647934, + 0.002939739730209112, + 0.005215875804424286, + 0.009022410027682781, + 0.007502589374780655, + 0.009022996760904789, + 0.01040284801274538, + 0.00931478850543499, + 0.006541538517922163, + 0.003715049009770155, + 0.028683992102742195, + 0.02951599285006523, + 0.020773891359567642, + 0.019987814128398895, + 0.03759951889514923, + 0.022005317732691765, + 0.020140226930379868, + 0.014124341309070587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.03812828287482262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23148779571056366, + 0.16743846237659454, + 0.11289241909980774, + 0.09774939715862274, + 0.03474314138293266, + 0.06128291040658951, + 0.02598787657916546, + 0.05697651207447052, + 0.1090502142906189, + 0.10239128023386002 + ], + "credit_map": [ + 0.01208218839019537, + 0.06924676895141602, + 0.01961512304842472, + 0.009977814741432667, + 0.04072059318423271, + 0.015387126244604588, + 0.04153584688901901, + 0.022922333329916, + 0.016423296183347702, + 0.04189170151948929, + 0.01569175161421299, + 0.009551974944770336, + 0.03623286262154579, + 0.020380914211273193, + 0.010368641465902328, + 0.016897330060601234, + 0.012577093206346035, + 0.014493521302938461, + 0.006946539506316185, + 0.015135619789361954, + 0.016248397529125214, + 0.021505206823349, + 0.011523786932229996, + 0.014462264254689217, + 0.0128327626734972, + 0.02155994065105915, + 0.0065003689378499985, + 0.015875648707151413, + 0.014127340167760849, + 0.008950464427471161, + 0.007625969592481852, + 0.010276906192302704, + 0.004572458565235138, + 0.007033620495349169, + 0.0030439826659858227, + 0.005041579715907574, + 0.003835191484540701, + 0.0038940536323934793, + 0.003072533756494522, + 0.004249718971550465, + 0.009306710213422775, + 0.00794159434735775, + 0.006618526764214039, + 0.00522535340860486, + 0.010184211656451225, + 0.007442844100296497, + 0.006667786743491888, + 0.007895881310105324, + 0.0032752971164882183, + 0.00318879634141922, + 0.002180777955800295, + 0.004336504731327295, + 0.003747296053916216, + 0.002673571929335594, + 0.0026218474376946688, + 0.003963787108659744, + 0.007794694975018501, + 0.006909551098942757, + 0.008432830683887005, + 0.005757838487625122, + 0.009323123842477798, + 0.006776923779398203, + 0.005869113840162754, + 0.006112437695264816, + 0.010463178157806396, + 0.011768982745707035, + 0.013942044228315353, + 0.018120067194104195, + 0.017111986875534058, + 0.017961248755455017, + 0.013578027486801147, + 0.0061046709306538105, + 0.016332440078258514, + 0.01842491328716278, + 0.012143668718636036, + 0.010493852198123932, + 0.01595856249332428, + 0.010258715599775314, + 0.010313878767192364, + 0.008465252816677094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005949762649834156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, majestic crane shot begins at a low angle, sweeping gracefully upward to reveal the breathtaking scale of a massive, multi-story greenhouse, rendered in a surreal and detailed anime-style composition. The structure, a soaring glass atrium and intricate white steel grid, creates a botanical cathedral. Inside, the air is filled with a soft, diffuse light. Vertical towers of luminous hydroponic plants rise from the polished concrete floor, their sleek frames supporting layers of lush, vibrant foliage, emitting a soft, ethereal teal glow. This cool interior glow contrasts beautifully with the warm golden rays of early morning sunlight filtering through the glass dome, casting long, dramatic rays of divine light that pierce the humid atmosphere. In classic Makoto Shinkai aesthetic, the entire scene is imbued with a sense of serene wonder, with countless glistening droplets of morning dew clinging to every leaf and glass with perfect surface tension. As the camera rises, these droplets catch the intersecting light sources, erupting in bright, sparkling anamorphic lens flares that dance across the frame, imbuing the futuristic yet natural environment with a palpable, shimmering energy.", + "seed": 15791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_2/index_8743_seed_15791_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027417022082954646, + "vlm_logprob": -0.027417022082954646, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.027417022082954646, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06297261267900467 + ], + "opposite_logprob_tokens": [ + -2.8129725456237793 + ], + "logprob_reward": -0.06297261267900467, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06297261267900467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09188695251941681, + 0.09793876856565475, + 0.03475265949964523, + 0.11041202396154404, + 0.06386805325746536, + 0.09101275354623795, + 0.09002106636762619, + 0.19974535703659058, + 0.1595044732093811, + 0.0608578622341156 + ], + "credit_map": [ + 0.010269690304994583, + 0.0178050696849823, + 0.00813772901892662, + 0.0067049781791865826, + 0.018615957349538803, + 0.01810184121131897, + 0.005698602646589279, + 0.006553084123879671, + 0.017322566360235214, + 0.009832575917243958, + 0.009524385444819927, + 0.006550578400492668, + 0.029299823567271233, + 0.014180946163833141, + 0.006320413667708635, + 0.004907485097646713, + 0.004354130942374468, + 0.0050133042968809605, + 0.003562389640137553, + 0.0024678052868694067, + 0.006793816573917866, + 0.007126916665583849, + 0.0026524390559643507, + 0.0027818637900054455, + 0.019307050853967667, + 0.016840411350131035, + 0.015715673565864563, + 0.012806518003344536, + 0.012038522399961948, + 0.018340259790420532, + 0.008066662587225437, + 0.007296934258192778, + 0.010988949798047543, + 0.007663305848836899, + 0.004520088899880648, + 0.007353400345891714, + 0.010138041339814663, + 0.005381162744015455, + 0.005099997855722904, + 0.012723117135465145, + 0.006298392545431852, + 0.00995256844907999, + 0.015962833538651466, + 0.012492755427956581, + 0.011399880051612854, + 0.005945520009845495, + 0.01684723049402237, + 0.012113584205508232, + 0.012526827864348888, + 0.014244184829294682, + 0.006473068613559008, + 0.007959920912981033, + 0.01971960812807083, + 0.010824989527463913, + 0.006989372428506613, + 0.011283100582659245, + 0.054357584565877914, + 0.05619444698095322, + 0.01749647594988346, + 0.02487749420106411, + 0.015349490568041801, + 0.008971686474978924, + 0.010955950245261192, + 0.011542238295078278, + 0.016034375876188278, + 0.02348703145980835, + 0.014736829325556755, + 0.012298720888793468, + 0.03653809055685997, + 0.019497154280543327, + 0.020739054307341576, + 0.016173243522644043, + 0.007095660082995892, + 0.00811043567955494, + 0.008210554718971252, + 0.010281183756887913, + 0.008282849565148354, + 0.00778495566919446, + 0.00491313124075532, + 0.006179099902510643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.06297261267900467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545830570161343 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545830570161343, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545830570161343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06560974568128586, + 0.13134780526161194, + 0.03281843662261963, + 0.038342300802469254, + 0.04939026013016701, + 0.027439936995506287, + 0.07190019637346268, + 0.10223519057035446, + 0.27382561564445496, + 0.20709045231342316 + ], + "credit_map": [ + 0.008967823348939419, + 0.011231942102313042, + 0.006986494176089764, + 0.004855107516050339, + 0.009381667710840702, + 0.013803242705762386, + 0.00421750545501709, + 0.006165972910821438, + 0.01944386214017868, + 0.015335778705775738, + 0.015425377525389194, + 0.006886368151754141, + 0.030725017189979553, + 0.023523155599832535, + 0.008783404715359211, + 0.011224853806197643, + 0.0036228890530765057, + 0.007309205364435911, + 0.004909049719572067, + 0.0030526036862283945, + 0.004938346333801746, + 0.004853188060224056, + 0.0021156221628189087, + 0.0020175338722765446, + 0.005984526127576828, + 0.004107848275452852, + 0.003910066559910774, + 0.00479925936087966, + 0.00641849497333169, + 0.007067455444484949, + 0.0032542156986892223, + 0.0028004352934658527, + 0.006775035988539457, + 0.0050636050291359425, + 0.005468617659062147, + 0.0071590314619243145, + 0.006150705739855766, + 0.004000887740403414, + 0.004947778303176165, + 0.009824605658650398, + 0.002355767646804452, + 0.002870341530069709, + 0.004318054299801588, + 0.003885324578732252, + 0.0037582905497401953, + 0.002257478889077902, + 0.004623089917004108, + 0.0033715926110744476, + 0.009404106065630913, + 0.010884263552725315, + 0.00697664450854063, + 0.008769935928285122, + 0.011247053742408752, + 0.00939215812832117, + 0.007094815373420715, + 0.008131228387355804, + 0.012917733751237392, + 0.01350799947977066, + 0.01596875861287117, + 0.017080694437026978, + 0.014865831471979618, + 0.0073425606824457645, + 0.009995924308896065, + 0.010555691085755825, + 0.024319564923644066, + 0.03752246871590614, + 0.02312728390097618, + 0.02726234681904316, + 0.06185474246740341, + 0.031334396451711655, + 0.04000139608979225, + 0.028403468430042267, + 0.031081244349479675, + 0.03439592942595482, + 0.022740699350833893, + 0.01919298619031906, + 0.04107928276062012, + 0.02488655596971512, + 0.0158572718501091, + 0.01785649172961712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.012545830570161343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637230679392815 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007637230679392815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637230679392815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09637051075696945, + 0.14407914876937866, + 0.08211955428123474, + 0.08770574629306793, + 0.08207646757364273, + 0.08430666476488113, + 0.06370813399553299, + 0.08320680260658264, + 0.1455416977405548, + 0.1308852583169937 + ], + "credit_map": [ + 0.010902095586061478, + 0.019666634500026703, + 0.00894702784717083, + 0.009670903906226158, + 0.017477497458457947, + 0.0153279360383749, + 0.006060394924134016, + 0.008318020962178707, + 0.02006533555686474, + 0.013131974264979362, + 0.026167938485741615, + 0.007037633564323187, + 0.03951025754213333, + 0.016566414386034012, + 0.009095191955566406, + 0.01250439789146185, + 0.013050291687250137, + 0.020015385001897812, + 0.010301725938916206, + 0.00782020203769207, + 0.01111013162881136, + 0.010657289996743202, + 0.005359293427318335, + 0.003805238287895918, + 0.011380690149962902, + 0.02247665449976921, + 0.007823960855603218, + 0.010086269117891788, + 0.011818795464932919, + 0.01067691296339035, + 0.00587023701518774, + 0.007572229951620102, + 0.008404959924519062, + 0.008761584758758545, + 0.009728348813951015, + 0.009930775500833988, + 0.013774961233139038, + 0.008474782109260559, + 0.00956042855978012, + 0.013440621085464954, + 0.005761729087680578, + 0.010366383008658886, + 0.01438889466226101, + 0.011422439478337765, + 0.012468216940760612, + 0.005574759095907211, + 0.014388230629265308, + 0.00993601605296135, + 0.008318641223013401, + 0.010122849605977535, + 0.0036041836719959974, + 0.003818626282736659, + 0.01511230319738388, + 0.009244422428309917, + 0.0067821647971868515, + 0.006704939063638449, + 0.018332786858081818, + 0.018840467557311058, + 0.007750249933451414, + 0.009632224217057228, + 0.0083217928186059, + 0.0064124418422579765, + 0.005638834089040756, + 0.008278013207018375, + 0.01806003227829933, + 0.02172601968050003, + 0.031041106209158897, + 0.02733100950717926, + 0.014491677284240723, + 0.011038469150662422, + 0.014122136868536472, + 0.007731261663138866, + 0.03573022782802582, + 0.03573450446128845, + 0.013896002434194088, + 0.009047442115843296, + 0.010642889887094498, + 0.007833761163055897, + 0.007309554610401392, + 0.010690873488783836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.007637230679392815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867362603545189 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.04867362603545189, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867362603545189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11353298276662827, + 0.14288468658924103, + 0.05272732302546501, + 0.05538275092840195, + 0.13709376752376556, + 0.05468573421239853, + 0.08846759051084518, + 0.1540614664554596, + 0.13191069662570953, + 0.06925296783447266 + ], + "credit_map": [ + 0.015296286903321743, + 0.026688003912568092, + 0.010134512558579445, + 0.00807795487344265, + 0.01980036124587059, + 0.017042912542819977, + 0.008352051489055157, + 0.00814089272171259, + 0.015375671908259392, + 0.020945560187101364, + 0.017233438789844513, + 0.012876505963504314, + 0.02019358240067959, + 0.026600852608680725, + 0.01256660558283329, + 0.017092473804950714, + 0.005579757038503885, + 0.0068127973936498165, + 0.008804588578641415, + 0.0041125440038740635, + 0.008137693628668785, + 0.007553138304501772, + 0.004578417167067528, + 0.007148385047912598, + 0.011558846570551395, + 0.012421056628227234, + 0.006491925101727247, + 0.00691859470680356, + 0.005299244541674852, + 0.005857263226062059, + 0.0036962348967790604, + 0.0031395854894071817, + 0.017541158944368362, + 0.013160115107893944, + 0.018653251230716705, + 0.018471218645572662, + 0.015212015248835087, + 0.010958279483020306, + 0.028181642293930054, + 0.01491608563810587, + 0.00431754719465971, + 0.005211000330746174, + 0.008461005985736847, + 0.008310213685035706, + 0.005616430193185806, + 0.0040112887509167194, + 0.01142947282642126, + 0.00732877291738987, + 0.011797377839684486, + 0.01572716422379017, + 0.006669031921774149, + 0.006537377834320068, + 0.019869286566972733, + 0.011832837015390396, + 0.008266755379736423, + 0.007767760660499334, + 0.01869017817080021, + 0.02063135616481304, + 0.025939567014575005, + 0.03338607773184776, + 0.016286246478557587, + 0.010347649455070496, + 0.013473455794155598, + 0.015306933782994747, + 0.01938338577747345, + 0.02255522832274437, + 0.02607034519314766, + 0.02195682004094124, + 0.01195624377578497, + 0.007605000864714384, + 0.015553015284240246, + 0.006830658297985792, + 0.007845781743526459, + 0.008712701499462128, + 0.010770177468657494, + 0.01108449138700962, + 0.00813867338001728, + 0.008077388629317284, + 0.007413353770971298, + 0.007210407871752977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.04867362603545189 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, majestic crane shot that begins at a low angle and sweeps upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a massive, multi-story greenhouse featuring a soaring glass atrium and intricate white steel grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vertical towers of luminous hydroponic plants visible rising from a polished concrete floor with layers of lush foliage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene display a Makoto Shinkai aesthetic with glistening morning dew droplets on leaves and glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sparkling anamorphic lens flares visible as the droplets catch the intersecting light sources?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05328832194209099, + 0.16549019515514374, + 0.035731978714466095, + 0.0993703156709671, + 0.05335262790322304, + 0.03743770718574524, + 0.058712270110845566, + 0.21576708555221558, + 0.13568301498889923, + 0.14516650140285492 + ], + "credit_map": [ + 0.00698884529992938, + 0.0090526407584548, + 0.005061598494648933, + 0.003666725242510438, + 0.009137965738773346, + 0.010478773154318333, + 0.004004685673862696, + 0.004897089675068855, + 0.018769728019833565, + 0.025367669761180878, + 0.018524019047617912, + 0.014983651228249073, + 0.021087555214762688, + 0.03232692554593086, + 0.014611546881496906, + 0.019819103181362152, + 0.004978653509169817, + 0.009259901009500027, + 0.00388707360252738, + 0.0023921679239720106, + 0.005086620803922415, + 0.006315304432064295, + 0.0020988881587982178, + 0.0017133713699877262, + 0.019242173060774803, + 0.03454843536019325, + 0.008199334144592285, + 0.009227977134287357, + 0.009456364437937737, + 0.0075803231447935104, + 0.004668241832405329, + 0.006447466555982828, + 0.005794701632112265, + 0.005372939631342888, + 0.007083038799464703, + 0.008077592588961124, + 0.005371508654206991, + 0.00622259546071291, + 0.00732957199215889, + 0.008100675418972969, + 0.002729275031015277, + 0.0035897393245249987, + 0.006126732565462589, + 0.006543537136167288, + 0.004554513841867447, + 0.0029913445468991995, + 0.00534667307510972, + 0.005555897019803524, + 0.007148794364184141, + 0.007152535021305084, + 0.006925568915903568, + 0.007652939297258854, + 0.010940726846456528, + 0.0053826007060706615, + 0.008317564614117146, + 0.0051915403455495834, + 0.04041758552193642, + 0.04443633183836937, + 0.024528352543711662, + 0.02941177412867546, + 0.019375989213585854, + 0.017955973744392395, + 0.014785263687372208, + 0.02485581673681736, + 0.019519299268722534, + 0.022857801988720894, + 0.016695987433195114, + 0.02476750873029232, + 0.014930842444300652, + 0.009258237667381763, + 0.01313339825719595, + 0.014519946649670601, + 0.03704460337758064, + 0.03735339269042015, + 0.015254765748977661, + 0.012833504006266594, + 0.010471838526427746, + 0.011921335943043232, + 0.008806989528238773, + 0.01148005947470665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.00525581045076251 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_tensor([4638, 4638])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_tensor([4638, 4638])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4fc119dd861594278d190d41bfa2aece350b8a94 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_tensor([4638, 4638])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot, held at water level, captures the moment a massive crocodile's snout deliberately breaks the glassy surface of a calm, murky river. Dappled sunlight, filtering through the invisible overhead canopy, glints on the damp, gnarled texture of its ancient, deep olive-green hide, highlighting every deep crack and armored scale. The camera remains steady as the prehistoric snout, partially covered by a layer of riverbed, remains virtually motionless, only its nostrils flaring almost imperceptibly. Just beneath the surface, the powerful leathery skin of its throat pulses in a subtle, rhythmic, and steadily deliberate pattern, its muscles contracting as it drinks. The surrounding river, dark and still, creates an intense, primal atmosphere, focusing all attention on the raw, latent power contained in this creature's quiet, watchful presence.", + "seed": 7831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7831_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02909576054662466, + "vlm_logprob": -0.02909576054662466, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02909576054662466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008089161477982998 + ], + "opposite_logprob_tokens": [ + -5.883089065551758 + ], + "logprob_reward": -0.008089161477982998, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008089161477982998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1811581254005432, + 0.07192075252532959, + 0.07304566353559494, + 0.05054706335067749, + 0.045299217104911804, + 0.2176421582698822, + 0.16771967709064484, + 0.10934733599424362, + 0.046354182064533234, + 0.03696583956480026 + ], + "credit_map": [ + 0.018580086529254913, + 0.02186373993754387, + 0.027428390458226204, + 0.02406909689307213, + 0.01712452620267868, + 0.02748563140630722, + 0.0247235968708992, + 0.01988304778933525, + 0.008032012730836868, + 0.009701961651444435, + 0.0071593341417610645, + 0.00527039123699069, + 0.010077354498207569, + 0.009745948947966099, + 0.011449972167611122, + 0.010483775287866592, + 0.006802260875701904, + 0.0064305877313017845, + 0.007877357304096222, + 0.010460647754371166, + 0.008926422335207462, + 0.008715336211025715, + 0.013056286610662937, + 0.010776754468679428, + 0.0045142751187086105, + 0.003958967514336109, + 0.004442492499947548, + 0.0056471736170351505, + 0.00932637695223093, + 0.006929631344974041, + 0.008610565215349197, + 0.007117574568837881, + 0.005084835924208164, + 0.004293214064091444, + 0.0054646702483296394, + 0.005582329351454973, + 0.004988039378076792, + 0.006410249508917332, + 0.006897941697388887, + 0.006577929947525263, + 0.024330582469701767, + 0.016006505116820335, + 0.02031097747385502, + 0.02218564972281456, + 0.028704583644866943, + 0.049889061599969864, + 0.021135056391358376, + 0.03507973626255989, + 0.012961248867213726, + 0.014608061872422695, + 0.01902070827782154, + 0.02293894998729229, + 0.015447353944182396, + 0.022611305117607117, + 0.03598816320300102, + 0.024143874645233154, + 0.008785451762378216, + 0.00787402130663395, + 0.012238473631441593, + 0.02131425030529499, + 0.011318936944007874, + 0.011428222991526127, + 0.019749941304326057, + 0.016638025641441345, + 0.004721594974398613, + 0.003462334629148245, + 0.006392563693225384, + 0.0063924421556293964, + 0.006923296954482794, + 0.005913775414228439, + 0.006691290996968746, + 0.005856877192854881, + 0.004655799362808466, + 0.004086241591721773, + 0.0048665851354599, + 0.004638485610485077, + 0.005067494697868824, + 0.004606584552675486, + 0.004308890551328659, + 0.004735750611871481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.008089161477982998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014230793341994286 + ], + "opposite_logprob_tokens": [ + -4.264230728149414 + ], + "logprob_reward": -0.014230793341994286, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014230793341994286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21675774455070496, + 0.1316683441400528, + 0.10995115339756012, + 0.09509754180908203, + 0.15179120004177094, + 0.0734780877828598, + 0.06221999228000641, + 0.06792306900024414, + 0.03122245892882347, + 0.05989043414592743 + ], + "credit_map": [ + 0.02196785807609558, + 0.031640514731407166, + 0.024368496611714363, + 0.028961168602108955, + 0.017349157482385635, + 0.030723823234438896, + 0.03435385599732399, + 0.02739286795258522, + 0.01250776369124651, + 0.012213259935379028, + 0.01224212534725666, + 0.014505242928862572, + 0.02586304023861885, + 0.01610155589878559, + 0.02126024104654789, + 0.016975123435258865, + 0.009777958504855633, + 0.008742752484977245, + 0.011572428978979588, + 0.020134251564741135, + 0.01407194510102272, + 0.010684153996407986, + 0.021270450204610825, + 0.013697206042706966, + 0.009704901836812496, + 0.008485371246933937, + 0.008557022549211979, + 0.011624517850577831, + 0.023712817579507828, + 0.01108220312744379, + 0.010283330455422401, + 0.011647376231849194, + 0.02207619696855545, + 0.017079461365938187, + 0.02052239514887333, + 0.0220219437032938, + 0.0161732267588377, + 0.018236855044960976, + 0.017668670043349266, + 0.018012460321187973, + 0.008834179490804672, + 0.005609648302197456, + 0.006242792587727308, + 0.006862462963908911, + 0.009779107756912708, + 0.016976654529571533, + 0.007092306390404701, + 0.01208093948662281, + 0.005744940601289272, + 0.006012375466525555, + 0.007479924708604813, + 0.006720465607941151, + 0.008540398441255093, + 0.013109588995575905, + 0.00613327044993639, + 0.008479024283587933, + 0.008014420978724957, + 0.00590856745839119, + 0.008987289853394032, + 0.008611674420535564, + 0.0059751784428954124, + 0.010308261029422283, + 0.01112410519272089, + 0.008993571624159813, + 0.00375218759290874, + 0.0030441484414041042, + 0.004788034129887819, + 0.004817915614694357, + 0.003605036763474345, + 0.0036770037841051817, + 0.003618379356339574, + 0.0039197527803480625, + 0.008508836850523949, + 0.009440664201974869, + 0.008491083979606628, + 0.008081579580903053, + 0.0066744782961905, + 0.00559223722666502, + 0.005921697244048119, + 0.007179860956966877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014230793341994286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819265007972717 + ], + "opposite_logprob_tokens": [ + -3.5298192501068115 + ], + "logprob_reward": -0.029819265007972717, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819265007972717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18354882299900055, + 0.14006781578063965, + 0.12481530755758286, + 0.03822668641805649, + 0.10897347331047058, + 0.1587037444114685, + 0.08400129526853561, + 0.0705239400267601, + 0.028395578265190125, + 0.06274335086345673 + ], + "credit_map": [ + 0.020389094948768616, + 0.019694464281201363, + 0.02387545071542263, + 0.02250799536705017, + 0.014962310902774334, + 0.030106179416179657, + 0.02933567576110363, + 0.022677641361951828, + 0.01629364863038063, + 0.016663920134305954, + 0.015945808961987495, + 0.011514290235936642, + 0.015732206404209137, + 0.02223554067313671, + 0.02243168279528618, + 0.019250700250267982, + 0.012035503052175045, + 0.012868483550846577, + 0.011283772997558117, + 0.01586613804101944, + 0.012812496162950993, + 0.01861780695617199, + 0.02551296353340149, + 0.015818150714039803, + 0.004080348182469606, + 0.003432514378800988, + 0.004310314543545246, + 0.0050156754441559315, + 0.005391787271946669, + 0.0045304191298782825, + 0.005465749651193619, + 0.005999878514558077, + 0.014892250299453735, + 0.012145229615271091, + 0.013763423077762127, + 0.012681068852543831, + 0.013113376684486866, + 0.014566580764949322, + 0.013918761163949966, + 0.013892781920731068, + 0.01689836196601391, + 0.012691907584667206, + 0.01357504166662693, + 0.01501469872891903, + 0.021911073476076126, + 0.03681182116270065, + 0.01494784839451313, + 0.026853008195757866, + 0.009183967486023903, + 0.006789573933929205, + 0.008324604481458664, + 0.009847612120211124, + 0.011909902095794678, + 0.01346602849662304, + 0.012442225590348244, + 0.012037372216582298, + 0.004827785771340132, + 0.005810323171317577, + 0.007951238192617893, + 0.010658688843250275, + 0.008475287817418575, + 0.008199149742722511, + 0.01400339137762785, + 0.010598067194223404, + 0.0028176310006529093, + 0.0024629130493849516, + 0.004646838176995516, + 0.004119251389056444, + 0.0025355189573019743, + 0.003531715366989374, + 0.004226289223879576, + 0.004055421333760023, + 0.007432229351252317, + 0.008249487727880478, + 0.007359985262155533, + 0.007024852558970451, + 0.011182872578501701, + 0.007404453586786985, + 0.006642599124461412, + 0.007446876261383295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029819265007972717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303097113966942 + ], + "opposite_logprob_tokens": [ + -3.7733030319213867 + ], + "logprob_reward": -0.023303097113966942, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303097113966942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1206311285495758, + 0.13974609971046448, + 0.2226620614528656, + 0.1438085436820984, + 0.0676644816994667, + 0.033402420580387115, + 0.044513460248708725, + 0.09473787993192673, + 0.04618336632847786, + 0.08665049821138382 + ], + "credit_map": [ + 0.011812332086265087, + 0.01159159280359745, + 0.01641303114593029, + 0.018513953313231468, + 0.010303226299583912, + 0.016904043033719063, + 0.021393703296780586, + 0.013699249364435673, + 0.01614028960466385, + 0.011601694859564304, + 0.010250954888761044, + 0.009866065345704556, + 0.027863644063472748, + 0.02445836178958416, + 0.020954489707946777, + 0.01861061155796051, + 0.024296050891280174, + 0.019500892609357834, + 0.02389155700802803, + 0.03201650083065033, + 0.03456783667206764, + 0.02339252270758152, + 0.0365833155810833, + 0.028413400053977966, + 0.012613417580723763, + 0.011722446419298649, + 0.014976311475038528, + 0.029291285201907158, + 0.015910234302282333, + 0.01602671481668949, + 0.025443051010370255, + 0.017825093120336533, + 0.009412758983671665, + 0.006763352081179619, + 0.006404309067875147, + 0.00778427068144083, + 0.007489115931093693, + 0.011031722649931908, + 0.008860749192535877, + 0.009918207302689552, + 0.004757375456392765, + 0.004210156854242086, + 0.0036469611804932356, + 0.002938386984169483, + 0.003313991939648986, + 0.0050116367638111115, + 0.004728919826447964, + 0.0047949934378266335, + 0.00555158406496048, + 0.004517121706157923, + 0.004873581696301699, + 0.0052278414368629456, + 0.005694881081581116, + 0.006884761620312929, + 0.006033316720277071, + 0.005730377044528723, + 0.007938419468700886, + 0.006933868397027254, + 0.010342366062104702, + 0.015210309065878391, + 0.009791667573153973, + 0.011762638576328754, + 0.01757282018661499, + 0.015185798518359661, + 0.004939110018312931, + 0.0045782714150846004, + 0.006633873097598553, + 0.006335758604109287, + 0.005958412308245897, + 0.006925982888787985, + 0.004504918586462736, + 0.0063070449978113174, + 0.01105315238237381, + 0.009275681339204311, + 0.013911676593124866, + 0.00899768341332674, + 0.010947532020509243, + 0.012548204511404037, + 0.009010547772049904, + 0.010906025767326355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.023303097113966942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003648579120636 + ], + "opposite_logprob_tokens": [ + -2.6950364112854004 + ], + "logprob_reward": -0.07003648579120636, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003648579120636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2749902904033661, + 0.31685614585876465, + 0.06460223346948624, + 0.02049669437110424, + 0.053072985261678696, + 0.024593325331807137, + 0.07022073119878769, + 0.13518281280994415, + 0.01070411130785942, + 0.029280656948685646 + ], + "credit_map": [ + 0.028897585347294807, + 0.038031816482543945, + 0.028553619980812073, + 0.02916024439036846, + 0.03003159910440445, + 0.035839807242155075, + 0.04955340549349785, + 0.03492222726345062, + 0.042650382965803146, + 0.029553940519690514, + 0.02310292236506939, + 0.0229302067309618, + 0.04575539007782936, + 0.058527007699012756, + 0.04910833761096001, + 0.045227959752082825, + 0.007766722235828638, + 0.0046644373796880245, + 0.005438316613435745, + 0.008048384450376034, + 0.004915557336062193, + 0.007772279437631369, + 0.01620856672525406, + 0.009787961840629578, + 0.0023005129769444466, + 0.0018946659984067082, + 0.0014858900103718042, + 0.0029106633737683296, + 0.003247788641601801, + 0.003010906046256423, + 0.0028543686494231224, + 0.002791897626593709, + 0.006232128478586674, + 0.005372676998376846, + 0.00690707191824913, + 0.006138029042631388, + 0.006272302009165287, + 0.008157743141055107, + 0.005599957425147295, + 0.008393073454499245, + 0.0031999009661376476, + 0.002112841233611107, + 0.003272205125540495, + 0.002744134049862623, + 0.0029803419020026922, + 0.004022402223199606, + 0.0029333282727748156, + 0.003328173654153943, + 0.00847942940890789, + 0.004071881528943777, + 0.0067412396892905235, + 0.009897017851471901, + 0.004802933428436518, + 0.008368403650820255, + 0.016357582062482834, + 0.011502246372401714, + 0.013606836088001728, + 0.009189554490149021, + 0.015187783166766167, + 0.017465882003307343, + 0.011749335564672947, + 0.020230431109666824, + 0.02789108268916607, + 0.019861919805407524, + 0.0012899014400318265, + 0.0007607918814755976, + 0.001151638338342309, + 0.0013881521299481392, + 0.0013184056151658297, + 0.0019767440389841795, + 0.0012179258046671748, + 0.00160055176820606, + 0.0028632055036723614, + 0.003105823416262865, + 0.0033645115327090025, + 0.004083622246980667, + 0.0031298722606152296, + 0.005551930516958237, + 0.002336702076718211, + 0.004844989627599716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07003648579120636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, held at water level, captures the moment a massive crocodile's snout deliberately breaks the glassy surface of a calm, murky river. Dappled sunlight, filtering through the invisible overhead canopy, glints on the damp, gnarled texture of its ancient, deep olive-green hide, highlighting every deep crack and armored scale. The camera remains steady as the prehistoric snout, partially covered by a layer of riverbed, remains virtually motionless, only its nostrils flaring almost imperceptibly. Just beneath the surface, the powerful leathery skin of its throat pulses in a subtle, rhythmic, and steadily deliberate pattern, its muscles contracting as it drinks. The surrounding river, dark and still, creates an intense, primal atmosphere, focusing all attention on the raw, latent power contained in this creature's quiet, watchful presence.", + "seed": 7832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7832_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023822648636996745, + "vlm_logprob": -0.023822648636996745, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.023822648636996745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011732857674360275 + ], + "opposite_logprob_tokens": [ + -4.886733055114746 + ], + "logprob_reward": -0.011732857674360275, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011732857674360275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12185429781675339, + 0.17219479382038116, + 0.2103038728237152, + 0.12869472801685333, + 0.1837415099143982, + 0.01953333057463169, + 0.03882059454917908, + 0.03437710180878639, + 0.05158938094973564, + 0.038890376687049866 + ], + "credit_map": [ + 0.009049322456121445, + 0.012461395002901554, + 0.013674018904566765, + 0.01682858355343342, + 0.016477664932608604, + 0.01586449146270752, + 0.02452501468360424, + 0.01297380868345499, + 0.013140367344021797, + 0.014429689384996891, + 0.021219946444034576, + 0.02417747862637043, + 0.02895800583064556, + 0.015121317468583584, + 0.03893769532442093, + 0.01621030643582344, + 0.01833225041627884, + 0.020282167941331863, + 0.024240544065833092, + 0.027824660763144493, + 0.030237331986427307, + 0.022637533023953438, + 0.04098870977759361, + 0.02576066367328167, + 0.004619404207915068, + 0.005024442449212074, + 0.024538572877645493, + 0.03594454005360603, + 0.014947276562452316, + 0.008672275580465794, + 0.024197416380047798, + 0.010750802233815193, + 0.003233762225136161, + 0.004854745697230101, + 0.008252989500761032, + 0.07543831318616867, + 0.006926494650542736, + 0.005815824493765831, + 0.06957753747701645, + 0.009641841985285282, + 0.0012698208447545767, + 0.0016016345471143723, + 0.0032355161383748055, + 0.0035586401354521513, + 0.0030349146109074354, + 0.0017953763017430902, + 0.0032063068356364965, + 0.0018311208114027977, + 0.001906693447381258, + 0.002075989032164216, + 0.008965235203504562, + 0.005840005353093147, + 0.0034422988537698984, + 0.0029694652184844017, + 0.007722659967839718, + 0.005898248869925737, + 0.0022004558704793453, + 0.002718839794397354, + 0.0049092997796833515, + 0.006097596138715744, + 0.00494479900225997, + 0.0034874563571065664, + 0.006896778475493193, + 0.003121875459328294, + 0.006241865921765566, + 0.003252235008403659, + 0.006398184224963188, + 0.009128060191869736, + 0.00635892478749156, + 0.00592840276658535, + 0.00823364220559597, + 0.006048065144568682, + 0.0012271367013454437, + 0.001320446957834065, + 0.004470474552363157, + 0.014232640154659748, + 0.002889030845835805, + 0.0015282176900655031, + 0.010562947951257229, + 0.0026594812516123056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.011732857674360275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008663804270327091 + ], + "opposite_logprob_tokens": [ + -4.758663654327393 + ], + "logprob_reward": -0.008663804270327091, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008663804270327091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06233403459191322, + 0.045638225972652435, + 0.03747912123799324, + 0.05873043090105057, + 0.6826198697090149, + 0.018337517976760864, + 0.035997457802295685, + 0.022269491106271744, + 0.019752580672502518, + 0.01684122160077095 + ], + "credit_map": [ + 0.004978823475539684, + 0.0052912263199687, + 0.010075409896671772, + 0.010185611434280872, + 0.007192923221737146, + 0.006204487755894661, + 0.012153241783380508, + 0.00625231908634305, + 0.003121207235381007, + 0.003531359601765871, + 0.00562336016446352, + 0.010418681427836418, + 0.0058920662850141525, + 0.004201965872198343, + 0.008708641864359379, + 0.004140948876738548, + 0.003220090875402093, + 0.003350997343659401, + 0.004590848460793495, + 0.005220083519816399, + 0.005817072000354528, + 0.0030894014053046703, + 0.007333898451179266, + 0.004856731276959181, + 0.0030709446873515844, + 0.0028249279130250216, + 0.008358638733625412, + 0.010235493071377277, + 0.008975919336080551, + 0.005774204153567553, + 0.01408473588526249, + 0.005405571311712265, + 0.006917304825037718, + 0.01607048325240612, + 0.01661364547908306, + 0.29371845722198486, + 0.01905812695622444, + 0.012308957986533642, + 0.2836577892303467, + 0.03427516296505928, + 0.0016010012477636337, + 0.0011235801503062248, + 0.0023902428802102804, + 0.0027592945843935013, + 0.0022902379278093576, + 0.001893057837150991, + 0.0044940682128071785, + 0.0017860373482108116, + 0.001620803028345108, + 0.002211775630712509, + 0.006948889698833227, + 0.008887537755072117, + 0.0026897310744971037, + 0.0026630316860973835, + 0.0049652280285954475, + 0.006010464858263731, + 0.0015486190095543861, + 0.0017616066616028547, + 0.0034247534349560738, + 0.003060214687138796, + 0.0032514729537069798, + 0.001927634933963418, + 0.004304274916648865, + 0.002990917768329382, + 0.0015850102063268423, + 0.0012086430797353387, + 0.0032449348364025354, + 0.0037517945747822523, + 0.0017163795419037342, + 0.002062401967123151, + 0.0036464459262788296, + 0.002536972751840949, + 0.0010057067265734076, + 0.0007785209454596043, + 0.002056099707260728, + 0.005492174997925758, + 0.001134765218012035, + 0.0013016010634601116, + 0.003782153595238924, + 0.001290200394578278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.008663804270327091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078807413578033 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078807413578033, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078807413578033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1124982088804245, + 0.07868940383195877, + 0.07607709616422653, + 0.0419590100646019, + 0.2547427713871002, + 0.026636777445673943, + 0.140397310256958, + 0.03118080645799637, + 0.08945150673389435, + 0.1483670175075531 + ], + "credit_map": [ + 0.009877583011984825, + 0.007852860726416111, + 0.015196731314063072, + 0.017979150637984276, + 0.011573923751711845, + 0.01479655783623457, + 0.023237673565745354, + 0.01198374293744564, + 0.0062667690217494965, + 0.006151421926915646, + 0.010913086123764515, + 0.014655254781246185, + 0.011148182675242424, + 0.009474709630012512, + 0.014245874248445034, + 0.005834112875163555, + 0.005770848598331213, + 0.005209995433688164, + 0.009059946984052658, + 0.011386096477508545, + 0.01013039518147707, + 0.008903715759515762, + 0.014956741593778133, + 0.010659360326826572, + 0.0019796991255134344, + 0.002268642419949174, + 0.006715364754199982, + 0.009374983608722687, + 0.00603858195245266, + 0.004014754667878151, + 0.006870690267533064, + 0.004696296527981758, + 0.004543997813016176, + 0.007169956341385841, + 0.009807812049984932, + 0.09862819314002991, + 0.012209867127239704, + 0.008404147811233997, + 0.10147390514612198, + 0.012504924088716507, + 0.0018857050454244018, + 0.0022535305470228195, + 0.0032455299515277147, + 0.003396499902009964, + 0.005729394499212503, + 0.0032107196748256683, + 0.004125491715967655, + 0.002789905993267894, + 0.0036314253229647875, + 0.005704191979020834, + 0.010849088430404663, + 0.05525544285774231, + 0.007248435635119677, + 0.006673574447631836, + 0.028542662039399147, + 0.02249250002205372, + 0.0016529106069356203, + 0.001535537769086659, + 0.005780585575848818, + 0.004562297370284796, + 0.002982701640576124, + 0.002235009567812085, + 0.007724676746875048, + 0.004707089625298977, + 0.009671882726252079, + 0.00585212092846632, + 0.007147022057324648, + 0.02017240971326828, + 0.012598824687302113, + 0.00675185676664114, + 0.016968276351690292, + 0.010289113968610764, + 0.006005547940731049, + 0.005731645505875349, + 0.014961222186684608, + 0.05289069935679436, + 0.011942117474973202, + 0.007332819979637861, + 0.04013330861926079, + 0.009369663894176483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.016078807413578033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590832456946373 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.020590832456946373, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590832456946373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11577664315700531, + 0.08201713860034943, + 0.04611630365252495, + 0.14443400502204895, + 0.2219688892364502, + 0.07668807357549667, + 0.21871261298656464, + 0.029824381694197655, + 0.03811727836728096, + 0.026344647631049156 + ], + "credit_map": [ + 0.009964382275938988, + 0.008566322736442089, + 0.011645070277154446, + 0.024113086983561516, + 0.010686662048101425, + 0.014692208729684353, + 0.023324575275182724, + 0.012784332036972046, + 0.006058442872017622, + 0.004001735709607601, + 0.014261226169764996, + 0.02223723568022251, + 0.008727790787816048, + 0.006807007826864719, + 0.012362179346382618, + 0.007561517413705587, + 0.0032960474491119385, + 0.0030414292123168707, + 0.005281346384435892, + 0.009094996377825737, + 0.005402794573456049, + 0.00495301466435194, + 0.008912674151360989, + 0.006134004332125187, + 0.0052304198034107685, + 0.0056022945791482925, + 0.027933955192565918, + 0.032890450209379196, + 0.015428495593369007, + 0.010264030657708645, + 0.034288953989744186, + 0.012795394286513329, + 0.008202598430216312, + 0.011389641091227531, + 0.017413407564163208, + 0.06429458409547806, + 0.017511408776044846, + 0.017078230157494545, + 0.0706372857093811, + 0.015441725961863995, + 0.003611707128584385, + 0.0035716721322387457, + 0.007434953935444355, + 0.015402157790958881, + 0.006189873442053795, + 0.006029399111866951, + 0.02758336439728737, + 0.006864944938570261, + 0.004741387441754341, + 0.007284244988113642, + 0.013428338803350925, + 0.09124533832073212, + 0.009235610254108906, + 0.008888176642358303, + 0.036150261759757996, + 0.047739241272211075, + 0.0017344814259558916, + 0.0017805668758228421, + 0.006589919328689575, + 0.003963935654610395, + 0.0027085861656814814, + 0.0021676854230463505, + 0.005179883912205696, + 0.005699321161955595, + 0.00261664018034935, + 0.0029364507645368576, + 0.004301895387470722, + 0.008078162558376789, + 0.005577266216278076, + 0.004916834644973278, + 0.005725197494029999, + 0.0039648329839110374, + 0.0012138421880081296, + 0.0011121585266664624, + 0.004066185560077429, + 0.00860660057514906, + 0.0014258059673011303, + 0.00192750315181911, + 0.005523940548300743, + 0.002468609483912587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.020590832456946373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204694136977196 + ], + "opposite_logprob_tokens": [ + -2.812047004699707 + ], + "logprob_reward": -0.06204694136977196, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204694136977196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07644019275903702, + 0.03094331920146942, + 0.015165310353040695, + 0.12769649922847748, + 0.1699417531490326, + 0.06098758056759834, + 0.21139433979988098, + 0.2552882134914398, + 0.029816050082445145, + 0.02232666127383709 + ], + "credit_map": [ + 0.007170191500335932, + 0.006372762843966484, + 0.007596634328365326, + 0.013316519558429718, + 0.01019810140132904, + 0.00933148991316557, + 0.01520485058426857, + 0.007249641697853804, + 0.002031185897067189, + 0.001614282839000225, + 0.005120604299008846, + 0.008295003324747086, + 0.003177989274263382, + 0.0031097952742129564, + 0.005246165674179792, + 0.002348296344280243, + 0.0011894063791260123, + 0.00125267228577286, + 0.0016729587223380804, + 0.0028135739266872406, + 0.002221745904535055, + 0.0018624304793775082, + 0.0025289789773523808, + 0.0016235454240813851, + 0.005794786382466555, + 0.005966359283775091, + 0.023765098303556442, + 0.027533581480383873, + 0.020336760208010674, + 0.01171910297125578, + 0.023093007504940033, + 0.009487809613347054, + 0.0051545691676437855, + 0.006513107102364302, + 0.009090597741305828, + 0.058676186949014664, + 0.00814888533204794, + 0.00759316748008132, + 0.06545765697956085, + 0.00930759310722351, + 0.005278709344565868, + 0.007049926090985537, + 0.00910835713148117, + 0.005134740844368935, + 0.009885765612125397, + 0.008188770152628422, + 0.010629560798406601, + 0.005711758509278297, + 0.004534794948995113, + 0.007489054463803768, + 0.0152538837864995, + 0.09162835776805878, + 0.005978933069854975, + 0.008119912818074226, + 0.034109074622392654, + 0.044280342757701874, + 0.012556713074445724, + 0.01405961811542511, + 0.04092182591557503, + 0.051552023738622665, + 0.03273029997944832, + 0.018325092270970345, + 0.06161462888121605, + 0.023528041318058968, + 0.002371307695284486, + 0.001625028788112104, + 0.004933485295623541, + 0.005257423967123032, + 0.00350372726097703, + 0.0027564067859202623, + 0.005470095202326775, + 0.003898578230291605, + 0.0011914677452296019, + 0.000997447525151074, + 0.0026866397820413113, + 0.007025396451354027, + 0.002151419408619404, + 0.0012987807858735323, + 0.005454399157315493, + 0.0015211133286356926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.06204694136977196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, held at water level, captures the moment a massive crocodile's snout deliberately breaks the glassy surface of a calm, murky river. Dappled sunlight, filtering through the invisible overhead canopy, glints on the damp, gnarled texture of its ancient, deep olive-green hide, highlighting every deep crack and armored scale. The camera remains steady as the prehistoric snout, partially covered by a layer of riverbed, remains virtually motionless, only its nostrils flaring almost imperceptibly. Just beneath the surface, the powerful leathery skin of its throat pulses in a subtle, rhythmic, and steadily deliberate pattern, its muscles contracting as it drinks. The surrounding river, dark and still, creates an intense, primal atmosphere, focusing all attention on the raw, latent power contained in this creature's quiet, watchful presence.", + "seed": 7833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7833_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032042961288243535, + "vlm_logprob": -0.032042961288243535, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.032042961288243535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028558073565363884 + ], + "opposite_logprob_tokens": [ + -3.7785580158233643 + ], + "logprob_reward": -0.028558073565363884, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028558073565363884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15308429300785065, + 0.02258007600903511, + 0.02298409305512905, + 0.04613738879561424, + 0.046018827706575394, + 0.1814143806695938, + 0.12938012182712555, + 0.12657366693019867, + 0.1754019558429718, + 0.09642517566680908 + ], + "credit_map": [ + 0.031103316694498062, + 0.016098910942673683, + 0.02007274329662323, + 0.012377609498798847, + 0.012853069230914116, + 0.018209069967269897, + 0.02533172070980072, + 0.01703786477446556, + 0.005387140903621912, + 0.001768421265296638, + 0.002766071353107691, + 0.001724532339721918, + 0.002215960994362831, + 0.002269901568070054, + 0.003918759059160948, + 0.0025292907375842333, + 0.005660546012222767, + 0.002291842130944133, + 0.002728079678490758, + 0.0016070784768089652, + 0.001668312819674611, + 0.002115032635629177, + 0.0038838121108710766, + 0.0030293890740722418, + 0.008738496340811253, + 0.0026168038602918386, + 0.009503897279500961, + 0.004920219536870718, + 0.002343264175578952, + 0.002964078215882182, + 0.0064770872704684734, + 0.008573549799621105, + 0.009400714188814163, + 0.0038867173716425896, + 0.005317283794283867, + 0.004188679624348879, + 0.002889168681576848, + 0.002533606020733714, + 0.00676690973341465, + 0.01103575062006712, + 0.05509112402796745, + 0.0038448437117040157, + 0.005251511465758085, + 0.003747048554942012, + 0.0033823896665126085, + 0.005279121920466423, + 0.027188939973711967, + 0.07762940973043442, + 0.033598240464925766, + 0.003785081207752228, + 0.006681626662611961, + 0.0029869305435568094, + 0.0028285905718803406, + 0.004853095859289169, + 0.022007830440998077, + 0.05263873562216759, + 0.03846723958849907, + 0.004076644312590361, + 0.006109634879976511, + 0.004850789438933134, + 0.005793484393507242, + 0.006485342979431152, + 0.04684838652610779, + 0.013942157849669456, + 0.053017791360616684, + 0.0036977401468902826, + 0.004962716717272997, + 0.004427273292094469, + 0.004724004305899143, + 0.007184810936450958, + 0.07115037739276886, + 0.026237251237034798, + 0.02964429184794426, + 0.002562900772318244, + 0.0031053125858306885, + 0.0030239231418818235, + 0.0021071704104542732, + 0.003568812971934676, + 0.013572409749031067, + 0.03884035721421242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.028558073565363884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006766383070498705 + ], + "opposite_logprob_tokens": [ + -5.006766319274902 + ], + "logprob_reward": -0.006766383070498705, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006766383070498705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06108566001057625, + 0.02418876811861992, + 0.02276674099266529, + 0.017832404002547264, + 0.018297793343663216, + 0.1965722143650055, + 0.30171728134155273, + 0.1420922726392746, + 0.06796010583639145, + 0.14748670160770416 + ], + "credit_map": [ + 0.012841226533055305, + 0.005603239871561527, + 0.006028167903423309, + 0.00543806841596961, + 0.00484915915876627, + 0.009298732504248619, + 0.009996217675507069, + 0.007030848413705826, + 0.004719160031527281, + 0.002067544497549534, + 0.0034854982513934374, + 0.0017681923927739263, + 0.002319391118362546, + 0.0029623783193528652, + 0.004087780602276325, + 0.002778823720291257, + 0.004614365752786398, + 0.0021522215101867914, + 0.0023237597197294235, + 0.0013997993664816022, + 0.0025486727245151997, + 0.0037375586107373238, + 0.003448663977906108, + 0.00254169967956841, + 0.0038004606030881405, + 0.0008594804094173014, + 0.0016793862450867891, + 0.001232824521139264, + 0.0014964818255975842, + 0.0017422015080228448, + 0.0026496106293052435, + 0.004371959716081619, + 0.00384167255833745, + 0.0014569518389180303, + 0.001733925542794168, + 0.00148850679397583, + 0.0013971657026559114, + 0.0026373472064733505, + 0.0027908096089959145, + 0.0029514150228351355, + 0.06330752372741699, + 0.004951020237058401, + 0.00668959878385067, + 0.004801188595592976, + 0.004002499394118786, + 0.00669087702408433, + 0.025110121816396713, + 0.08101939409971237, + 0.08952094614505768, + 0.00599087867885828, + 0.005200743209570646, + 0.00453220447525382, + 0.004114964511245489, + 0.00966547429561615, + 0.05019973963499069, + 0.132492333650589, + 0.037958044558763504, + 0.002876747865229845, + 0.0027451268397271633, + 0.0038393468130379915, + 0.004204899538308382, + 0.00578061630949378, + 0.06047235429286957, + 0.024215135723352432, + 0.017515577375888824, + 0.001369415782392025, + 0.0017283973284065723, + 0.0017337872413918376, + 0.0018464629538357258, + 0.0026691285893321037, + 0.028672875836491585, + 0.012424462474882603, + 0.04448791965842247, + 0.0035948646254837513, + 0.005476807244122028, + 0.005332790780812502, + 0.003439118154346943, + 0.0057663461193442345, + 0.021404698491096497, + 0.057984158396720886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.006766383070498705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486757829785347 + ], + "opposite_logprob_tokens": [ + -3.048675775527954 + ], + "logprob_reward": -0.0486757829785347, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486757829785347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12151376157999039, + 0.06422461569309235, + 0.045390479266643524, + 0.0140529228374362, + 0.01625564694404602, + 0.31088462471961975, + 0.141346275806427, + 0.11785940825939178, + 0.031641583889722824, + 0.13683076202869415 + ], + "credit_map": [ + 0.029181763529777527, + 0.013446558266878128, + 0.01296195201575756, + 0.009778205305337906, + 0.010036223568022251, + 0.010628669522702694, + 0.020830459892749786, + 0.014649936929345131, + 0.014093656092882156, + 0.006565071176737547, + 0.010865649208426476, + 0.004529344849288464, + 0.0036815006751567125, + 0.0032612797804176807, + 0.011688758619129658, + 0.009539355523884296, + 0.01241274643689394, + 0.005441954359412193, + 0.00423082010820508, + 0.0033594975247979164, + 0.00259296759031713, + 0.0024202736094594, + 0.007825633510947227, + 0.00710658123716712, + 0.0029729052912443876, + 0.0010233551729470491, + 0.002342639258131385, + 0.0014508605236187577, + 0.0010314760729670525, + 0.0009813610231503844, + 0.0021011349745094776, + 0.0021491905208677053, + 0.0029206564649939537, + 0.0013974419562146068, + 0.001546798157505691, + 0.0013825756032019854, + 0.000798062130343169, + 0.0012745388085022569, + 0.002690250752493739, + 0.004245322663336992, + 0.09851978719234467, + 0.006173315923660994, + 0.005840859841555357, + 0.0056021129712462425, + 0.004519541282206774, + 0.00873641949146986, + 0.0455344058573246, + 0.13595817983150482, + 0.0394706092774868, + 0.003188439179211855, + 0.004627486225217581, + 0.0032936371862888336, + 0.0026095376815646887, + 0.004357011988759041, + 0.023569229990243912, + 0.06023033335804939, + 0.013784883543848991, + 0.0019801384769380093, + 0.0033458129037171602, + 0.0038677325937896967, + 0.0035090509336441755, + 0.0035905414260923862, + 0.050739586353302, + 0.03704166039824486, + 0.010151797905564308, + 0.0012347933370620012, + 0.0028813236858695745, + 0.002098443452268839, + 0.0008654382545500994, + 0.0016722441650927067, + 0.008662115782499313, + 0.004075426142662764, + 0.04286346957087517, + 0.003577571362257004, + 0.003964729141443968, + 0.006553132552653551, + 0.002931163879111409, + 0.006459936499595642, + 0.016715727746486664, + 0.05376502871513367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.0486757829785347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811141476035118 + ], + "opposite_logprob_tokens": [ + -3.288111448287964 + ], + "logprob_reward": -0.03811141476035118, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811141476035118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06251920759677887, + 0.0226017814129591, + 0.024392060935497284, + 0.018379393965005875, + 0.021476486697793007, + 0.0947781652212143, + 0.3263070583343506, + 0.14857575297355652, + 0.0811135545372963, + 0.19985653460025787 + ], + "credit_map": [ + 0.014904097653925419, + 0.006473194342106581, + 0.0072417957708239555, + 0.004347133915871382, + 0.004470283165574074, + 0.0061919051222503185, + 0.01082821935415268, + 0.008062576875090599, + 0.005633366294205189, + 0.002288172720000148, + 0.003568389918655157, + 0.001852343208156526, + 0.001045468496158719, + 0.0014844242250546813, + 0.002829104894772172, + 0.0039005109574645758, + 0.005400607828050852, + 0.0033535759430378675, + 0.0029651657678186893, + 0.001872124383226037, + 0.0016990075819194317, + 0.001821759156882763, + 0.003538520308211446, + 0.003741299966350198, + 0.004052550997585058, + 0.0016978790517896414, + 0.0022731327917426825, + 0.0023196518886834383, + 0.0013469301629811525, + 0.0011829747818410397, + 0.00245718308724463, + 0.0030490909703075886, + 0.004223082680255175, + 0.001782068400643766, + 0.0020768276881426573, + 0.0017786248354241252, + 0.0012855005916208029, + 0.0013877343153581023, + 0.0035200698766857386, + 0.005422578193247318, + 0.020437661558389664, + 0.008499795570969582, + 0.010238789953291416, + 0.006108890287578106, + 0.009256167337298393, + 0.008301706984639168, + 0.018332716077566147, + 0.013602431863546371, + 0.09569218754768372, + 0.005962573457509279, + 0.00577615387737751, + 0.0045291525311768055, + 0.004247674252837896, + 0.01030014455318451, + 0.05580717697739601, + 0.14399199187755585, + 0.03200177475810051, + 0.0027080285362899303, + 0.0036794166080653667, + 0.004247082397341728, + 0.004556703846901655, + 0.005611141212284565, + 0.06449412554502487, + 0.03127747401595116, + 0.014517532661557198, + 0.0021731341257691383, + 0.0037994063459336758, + 0.0035240089055150747, + 0.0030641553457826376, + 0.0038104134146124125, + 0.030652128159999847, + 0.019572772085666656, + 0.07022811472415924, + 0.0054033296182751656, + 0.006502293050289154, + 0.006402817089110613, + 0.004376500844955444, + 0.008920320309698582, + 0.02154766023159027, + 0.07647550851106644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.03811141476035118 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810315206646919 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810315206646919, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810315206646919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038983624428510666, + 0.029283955693244934, + 0.012961660511791706, + 0.018232129514217377, + 0.04889099299907684, + 0.150032639503479, + 0.5028705596923828, + 0.018937798216938972, + 0.04725320637226105, + 0.13255351781845093 + ], + "credit_map": [ + 0.00614596251398325, + 0.0030837052036076784, + 0.005594945512712002, + 0.005382684525102377, + 0.0027094674296677113, + 0.004348013084381819, + 0.006517627742141485, + 0.005201215855777264, + 0.005324590019881725, + 0.001903223223052919, + 0.00554560124874115, + 0.0022034673020243645, + 0.002336690202355385, + 0.00266777235083282, + 0.005601515062153339, + 0.003701096633449197, + 0.0024073750246316195, + 0.0010033876169472933, + 0.001561689656227827, + 0.00132778100669384, + 0.0009494315600022674, + 0.0018008718034252524, + 0.001955764601007104, + 0.001955359010025859, + 0.005511049646884203, + 0.0007167744333855808, + 0.001385448151268065, + 0.0009358314564451575, + 0.0005260763573460281, + 0.0008838321664370596, + 0.0023307998199015856, + 0.00594231765717268, + 0.010564425028860569, + 0.0036922278814017773, + 0.005809026770293713, + 0.0054463474079966545, + 0.0028100027702748775, + 0.002661320613697171, + 0.007225522771477699, + 0.010682119987905025, + 0.04430627450346947, + 0.003086154581978917, + 0.004021079745143652, + 0.0038760025054216385, + 0.0029264786280691624, + 0.004918332677334547, + 0.023862246423959732, + 0.06303606927394867, + 0.14557446539402008, + 0.009709774516522884, + 0.008492599241435528, + 0.007114019710570574, + 0.00664736982434988, + 0.015881406143307686, + 0.08702640235424042, + 0.2224244922399521, + 0.0073212310671806335, + 0.0006564979557879269, + 0.0010527875274419785, + 0.0007918060291558504, + 0.0004525338299572468, + 0.0007953330641612411, + 0.004122861195355654, + 0.0037447481881827116, + 0.0156483706086874, + 0.0009703382384032011, + 0.0011983992299064994, + 0.0014531239867210388, + 0.0012805452570319176, + 0.0020014732144773006, + 0.0187215693295002, + 0.005979386158287525, + 0.04089319705963135, + 0.0027713344898074865, + 0.003666421864181757, + 0.0041161454282701015, + 0.002393599832430482, + 0.004880903754383326, + 0.01989460363984108, + 0.05393730103969574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.03810315206646919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, held at water level, captures the moment a massive crocodile's snout deliberately breaks the glassy surface of a calm, murky river. Dappled sunlight, filtering through the invisible overhead canopy, glints on the damp, gnarled texture of its ancient, deep olive-green hide, highlighting every deep crack and armored scale. The camera remains steady as the prehistoric snout, partially covered by a layer of riverbed, remains virtually motionless, only its nostrils flaring almost imperceptibly. Just beneath the surface, the powerful leathery skin of its throat pulses in a subtle, rhythmic, and steadily deliberate pattern, its muscles contracting as it drinks. The surrounding river, dark and still, creates an intense, primal atmosphere, focusing all attention on the raw, latent power contained in this creature's quiet, watchful presence.", + "seed": 7834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_3/index_4638_seed_7834_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030423777177929878, + "vlm_logprob": -0.030423777177929878, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.030423777177929878, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01742752455174923 + ], + "opposite_logprob_tokens": [ + -4.642427444458008 + ], + "logprob_reward": -0.01742752455174923, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01742752455174923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06472377479076385, + 0.06566113978624344, + 0.16848041117191315, + 0.18371696770191193, + 0.012139223515987396, + 0.014451127499341965, + 0.04952927678823471, + 0.05519343912601471, + 0.31134918332099915, + 0.07475543767213821 + ], + "credit_map": [ + 0.004287300165742636, + 0.007479244377464056, + 0.003948583267629147, + 0.003897362854331732, + 0.01938619092106819, + 0.013362144120037556, + 0.004948067478835583, + 0.007414878346025944, + 0.005843815393745899, + 0.01041490025818348, + 0.006334803067147732, + 0.006702460814267397, + 0.011586776934564114, + 0.01066158339381218, + 0.006928241811692715, + 0.007188551127910614, + 0.012976312078535557, + 0.019150273874402046, + 0.01862080581486225, + 0.017284899950027466, + 0.027981145307421684, + 0.02333586849272251, + 0.026871776208281517, + 0.022259319201111794, + 0.02222704142332077, + 0.023518161848187447, + 0.0257912278175354, + 0.015716712921857834, + 0.03694257512688637, + 0.028226245194673538, + 0.015425855293869972, + 0.015869149938225746, + 0.001090449164621532, + 0.0012795277871191502, + 0.0016515377210453153, + 0.0009973912965506315, + 0.00199459632858634, + 0.0025950574781745672, + 0.0011315117590129375, + 0.00139915244653821, + 0.0011426914716139436, + 0.0016084153903648257, + 0.002071226481348276, + 0.001365039381198585, + 0.0025397443678230047, + 0.002129988046362996, + 0.0017446078127250075, + 0.0018494153628125787, + 0.0033370796591043472, + 0.005211889743804932, + 0.004517716821283102, + 0.004663798492401838, + 0.00710010901093483, + 0.0078064315021038055, + 0.010639340616762638, + 0.00625291233882308, + 0.005272928159683943, + 0.005731008946895599, + 0.0066047064028680325, + 0.005694627296179533, + 0.008900342509150505, + 0.007278298493474722, + 0.007821457460522652, + 0.007890067063272, + 0.03313019499182701, + 0.041326455771923065, + 0.0302495826035738, + 0.02205631323158741, + 0.06468383967876434, + 0.06560730189085007, + 0.021171536296606064, + 0.03312396630644798, + 0.007640224415808916, + 0.012388844043016434, + 0.011033976450562477, + 0.010359153151512146, + 0.008117531426250935, + 0.010405574925243855, + 0.009247309528291225, + 0.005562822800129652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.01742752455174923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007657578215003014 + ], + "opposite_logprob_tokens": [ + -4.882657527923584 + ], + "logprob_reward": -0.007657578215003014, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007657578215003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10473469644784927, + 0.13103777170181274, + 0.17836520075798035, + 0.17539726197719574, + 0.040999509394168854, + 0.017118187621235847, + 0.021801555529236794, + 0.03313284367322922, + 0.1569443792104721, + 0.14046861231327057 + ], + "credit_map": [ + 0.010076940059661865, + 0.01026793010532856, + 0.013688544742763042, + 0.01104925386607647, + 0.018393639475107193, + 0.019087374210357666, + 0.00756572000682354, + 0.014605282805860043, + 0.011764832772314548, + 0.016148582100868225, + 0.016947634518146515, + 0.011135507375001907, + 0.025410428643226624, + 0.025963997468352318, + 0.01059387531131506, + 0.013072898611426353, + 0.01334663201123476, + 0.018012458458542824, + 0.020413106307387352, + 0.01876765675842762, + 0.03153681755065918, + 0.02665901929140091, + 0.025803424417972565, + 0.023826060816645622, + 0.01585777848958969, + 0.022716708481311798, + 0.020114731043577194, + 0.02704503946006298, + 0.025432651862502098, + 0.02213100716471672, + 0.022385116666555405, + 0.019714219495654106, + 0.0032946469727903605, + 0.00451986538246274, + 0.0050313775427639484, + 0.004299165681004524, + 0.006184019614011049, + 0.008007432334125042, + 0.004172913730144501, + 0.005490086507052183, + 0.0015592401614412665, + 0.0018049008212983608, + 0.0016730802599340677, + 0.0017964948201552033, + 0.0031676129437983036, + 0.0026257489807903767, + 0.002375603886321187, + 0.002115504117682576, + 0.0015140898758545518, + 0.001918391208164394, + 0.002842626301571727, + 0.0017115323571488261, + 0.003628350328654051, + 0.003064619842916727, + 0.003703923663124442, + 0.0034180183429270983, + 0.0031439680606126785, + 0.003556253621354699, + 0.0039092181250452995, + 0.0021138552110642195, + 0.006429499480873346, + 0.004913894459605217, + 0.004220578353852034, + 0.0048455712385475636, + 0.012825009413063526, + 0.018869955092668533, + 0.01767849177122116, + 0.011285435408353806, + 0.040176864713430405, + 0.026382125914096832, + 0.014554043300449848, + 0.01517243217676878, + 0.009823404252529144, + 0.017014293000102043, + 0.02015611343085766, + 0.02167144976556301, + 0.014514701440930367, + 0.02076764404773712, + 0.025385089218616486, + 0.011135896667838097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.007657578215003014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372057527303696 + ], + "opposite_logprob_tokens": [ + -3.4087204933166504 + ], + "logprob_reward": -0.03372057527303696, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372057527303696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07926606386899948, + 0.10384233295917511, + 0.3271593451499939, + 0.0692175105214119, + 0.03397110104560852, + 0.020344579592347145, + 0.015577510930597782, + 0.021659906953573227, + 0.24508143961429596, + 0.08388015627861023 + ], + "credit_map": [ + 0.006882578134536743, + 0.010165528394281864, + 0.006983764469623566, + 0.005071277264505625, + 0.02035858854651451, + 0.014084038324654102, + 0.006128400564193726, + 0.009591886773705482, + 0.009462272748351097, + 0.011763439513742924, + 0.013429205864667892, + 0.011765756644308567, + 0.02096349373459816, + 0.01708902046084404, + 0.00925406813621521, + 0.010115073062479496, + 0.02996637485921383, + 0.035766396671533585, + 0.035569995641708374, + 0.0396149568259716, + 0.04463612288236618, + 0.04321149364113808, + 0.0590779110789299, + 0.039316099137067795, + 0.006480156909674406, + 0.007689823396503925, + 0.009925887919962406, + 0.006010222714394331, + 0.014165223576128483, + 0.01072357315570116, + 0.00527686532586813, + 0.008945760317146778, + 0.0023041057866066694, + 0.003486537840217352, + 0.004798335488885641, + 0.0037081532645970583, + 0.006434980779886246, + 0.005838255863636732, + 0.00305207259953022, + 0.004348660819232464, + 0.0012626803945749998, + 0.002165150595828891, + 0.002427253872156143, + 0.001674498664215207, + 0.0036357794888317585, + 0.003172011114656925, + 0.0027954205870628357, + 0.003211785340681672, + 0.0009435679530724883, + 0.0023313683923333883, + 0.0017442333046346903, + 0.001627618563361466, + 0.0025951145216822624, + 0.002168813953176141, + 0.001952241058461368, + 0.0022145542316138744, + 0.00275603705085814, + 0.0027501718141138554, + 0.0026808485854417086, + 0.002239090856164694, + 0.0027998853474855423, + 0.0029247007332742214, + 0.0031072148121893406, + 0.002401959151029587, + 0.023829853162169456, + 0.03016219288110733, + 0.02765789069235325, + 0.017579365521669388, + 0.05162239819765091, + 0.052198369055986404, + 0.017448963597416878, + 0.024582412093877792, + 0.007663551717996597, + 0.01326246839016676, + 0.008987333625555038, + 0.008783315308392048, + 0.009596124291419983, + 0.013815712183713913, + 0.00937074888497591, + 0.012400904670357704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.03372057527303696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294946178793907 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294946178793907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294946178793907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09894035011529922, + 0.05018703266978264, + 0.1987396478652954, + 0.11488045752048492, + 0.0317360945045948, + 0.03891284391283989, + 0.04106956347823143, + 0.04154221713542938, + 0.30945929884910583, + 0.07453246414661407 + ], + "credit_map": [ + 0.007547159679234028, + 0.012456282041966915, + 0.012693230994045734, + 0.009423530660569668, + 0.017459891736507416, + 0.013375537469983101, + 0.010550363920629025, + 0.015434356406331062, + 0.004143334925174713, + 0.007532864809036255, + 0.00502141285687685, + 0.005579215940088034, + 0.009739446453750134, + 0.00715356832370162, + 0.004903341177850962, + 0.006113852374255657, + 0.015238065272569656, + 0.02035980112850666, + 0.020862268283963203, + 0.023926183581352234, + 0.03173087164759636, + 0.02632220834493637, + 0.0369412861764431, + 0.02335898019373417, + 0.012686890549957752, + 0.01371289137750864, + 0.015349434688687325, + 0.01537526398897171, + 0.018487364053726196, + 0.016066458076238632, + 0.012556876987218857, + 0.010645278729498386, + 0.002263536909595132, + 0.003961028065532446, + 0.004073033574968576, + 0.003802119754254818, + 0.005166311748325825, + 0.005197019316256046, + 0.003537526121363044, + 0.003735520411282778, + 0.0028202871326357126, + 0.004312220029532909, + 0.0033184350468218327, + 0.0028707250021398067, + 0.008135681971907616, + 0.00817564595490694, + 0.00520957401022315, + 0.004070276860147715, + 0.002656051889061928, + 0.005783210042864084, + 0.004911184310913086, + 0.004832472186535597, + 0.007152084726840258, + 0.00618336908519268, + 0.004871156066656113, + 0.004680038895457983, + 0.004014208447188139, + 0.005297965370118618, + 0.007012300658971071, + 0.003845615778118372, + 0.005923429504036903, + 0.005032480228692293, + 0.005019413772970438, + 0.005396808031946421, + 0.028820300474762917, + 0.03997199982404709, + 0.028599359095096588, + 0.019750472158193588, + 0.07169775664806366, + 0.06154590845108032, + 0.024924881756305695, + 0.034148626029491425, + 0.00607415521517396, + 0.007355303969234228, + 0.01292265485972166, + 0.011924498714506626, + 0.007146743591874838, + 0.01170033123344183, + 0.011258983984589577, + 0.006149795837700367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.023294946178793907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a massive crocodile's snout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crocodile's hide deep olive-green with visible deep cracks and armored scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crocodile's snout break the glassy surface of a calm, murky river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held at water level and remains steady throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is dappled sunlight visible filtering through an overhead canopy onto the crocodile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001826167106628 + ], + "opposite_logprob_tokens": [ + -2.6950182914733887 + ], + "logprob_reward": -0.07001826167106628, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001826167106628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061100609600543976, + 0.04545305296778679, + 0.21884849667549133, + 0.5023451447486877, + 0.028969764709472656, + 0.008173426613211632, + 0.016903992742300034, + 0.019612502306699753, + 0.019495517015457153, + 0.07909746468067169 + ], + "credit_map": [ + 0.005131884012371302, + 0.007686472497880459, + 0.006744177080690861, + 0.005898819770663977, + 0.013239211402833462, + 0.010565849021077156, + 0.005853478796780109, + 0.005980712827295065, + 0.003995396196842194, + 0.005276367999613285, + 0.006450282875448465, + 0.0030708895064890385, + 0.011049674823880196, + 0.007616700138896704, + 0.0034630722366273403, + 0.004530670121312141, + 0.01954634115099907, + 0.022334378212690353, + 0.02525460533797741, + 0.025857042521238327, + 0.03190794959664345, + 0.02906464971601963, + 0.0373370386660099, + 0.02754649892449379, + 0.05284209921956062, + 0.06318595260381699, + 0.07136068493127823, + 0.06509189307689667, + 0.08067932724952698, + 0.06313442438840866, + 0.05366958677768707, + 0.052381210029125214, + 0.0017827226547524333, + 0.003607546677812934, + 0.002375283744186163, + 0.002814655192196369, + 0.006041881628334522, + 0.004790411330759525, + 0.0035636015236377716, + 0.003993664402514696, + 0.0005909799365326762, + 0.0009154475410468876, + 0.0009225686080753803, + 0.0006131306290626526, + 0.0016807577339932323, + 0.0012488889042288065, + 0.001034306362271309, + 0.001167346490547061, + 0.001207292778417468, + 0.0023647842463105917, + 0.001549741136841476, + 0.0016872151754796505, + 0.00234951707534492, + 0.0024778866209089756, + 0.003337982576340437, + 0.0019295724341645837, + 0.001593615161255002, + 0.0014734519645571709, + 0.0031336741521954536, + 0.001969088101759553, + 0.0030621555633842945, + 0.0025589424185454845, + 0.0027491834480315447, + 0.003072391962632537, + 0.002298889448866248, + 0.001996513456106186, + 0.0018256265902891755, + 0.0014582693111151457, + 0.003986669238656759, + 0.004736636765301228, + 0.0017726972000673413, + 0.0014202138409018517, + 0.0070344735868275166, + 0.005852802656590939, + 0.011413611471652985, + 0.011676052585244179, + 0.007249840069562197, + 0.016453899443149567, + 0.010130215436220169, + 0.009286572225391865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.07001826167106628 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_tensor([8403, 8403])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_tensor([8403, 8403])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..25a9c4e1128a34a40c953dad9b8e3aff6cdfc57d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_tensor([8403, 8403])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed, low-angle, wide-angle shot focuses on a downpour-soaked asphalt street in a cyberpunk city. Between towering skyscrapers, giant, translucent holographic advertisements flicker with shifting patterns, interplaying with pink and cyan neon signs, painting the slippery road surface into a shimmering mirror. The air is thick with damp vapor. A heavily modified off-road vehicle suddenly bursts into view, its matte black armored body covered in battle-worn scratches and mud stains. The blue LED strips on its underbody cast an eerie halo of light on the ground. The vehicle drifts with incredible speed and precision. The wide, heavily treaded off-road tires lose grip the moment they make contact, emitting a sharp, grinding sound, and the vehicle tilts violently under the immense force of inertia. At this moment, the high-speed spinning rear wheel suddenly rolled up the accumulated water on the road, exploding a huge fan-shaped water curtain that almost occupied half of the screen. Every splashing water droplet reflected a diamond-like bright and short-lived colorful light under the illumination of the surrounding chaotic light. The whole scene was full of movement and power. Against the backdrop of the cold rainy night and the dazzling technological light pollution, it created a tense, chaotic and visually impactful futuristic atmosphere.", + "seed": 8682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_8403_seed_8682_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024190970323979856, + "vlm_logprob": -0.024190970323979856, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.024190970323979856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047720976173877716 + ], + "opposite_logprob_tokens": [ + -3.1727209091186523 + ], + "logprob_reward": -0.047720976173877716, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.047720976173877716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15134777128696442, + 0.2695642113685608, + 0.12847313284873962, + 0.07485643774271011, + 0.03548334911465645, + 0.037388093769550323, + 0.18521648645401, + 0.044194091111421585, + 0.02210591360926628, + 0.051370542496442795 + ], + "credit_map": [ + 0.014867164194583893, + 0.017537694424390793, + 0.019198566675186157, + 0.015130635350942612, + 0.019045988097786903, + 0.01270472165197134, + 0.026575708761811256, + 0.02628728747367859, + 0.035333260893821716, + 0.040824778378009796, + 0.0258486270904541, + 0.04559530317783356, + 0.03760920464992523, + 0.01984451524913311, + 0.03430944308638573, + 0.030199069529771805, + 0.010823821648955345, + 0.008586264215409756, + 0.03769770637154579, + 0.010269111953675747, + 0.008409952744841576, + 0.016821594908833504, + 0.013889304362237453, + 0.021975373849272728, + 0.0067605688236653805, + 0.006621037609875202, + 0.015143954195082188, + 0.009764259681105614, + 0.00787126924842596, + 0.008305256254971027, + 0.0098948422819376, + 0.010495251044631004, + 0.004641071893274784, + 0.0026712410617619753, + 0.004852660465985537, + 0.006354406010359526, + 0.0038952564354985952, + 0.004545674193650484, + 0.005044087301939726, + 0.003478950820863247, + 0.003170279785990715, + 0.00800236314535141, + 0.007014610338956118, + 0.004616342484951019, + 0.0027118197176605463, + 0.0035102996043860912, + 0.00401091855019331, + 0.004351458046585321, + 0.022485019639134407, + 0.01473454013466835, + 0.03070491924881935, + 0.02150656469166279, + 0.017095664516091347, + 0.027749642729759216, + 0.030091261491179466, + 0.020848875865340233, + 0.004290539771318436, + 0.004713705275207758, + 0.0040249889716506, + 0.010559218004345894, + 0.0063010030426084995, + 0.00333287101238966, + 0.005629354622215033, + 0.005342413205653429, + 0.0030911790672689676, + 0.003982825670391321, + 0.002578514162451029, + 0.0025670193135738373, + 0.0019066992681473494, + 0.002030156785622239, + 0.0022675050422549248, + 0.003682016162201762, + 0.004122589249163866, + 0.006697840057313442, + 0.00492012407630682, + 0.0067868405021727085, + 0.004813060164451599, + 0.00520370714366436, + 0.009197601117193699, + 0.009628781117498875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.047720976173877716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963864270597696 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.005963864270597696, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963864270597696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07424965500831604, + 0.11090472340583801, + 0.09617616981267929, + 0.2030554562807083, + 0.07076793909072876, + 0.07196778059005737, + 0.15410509705543518, + 0.09682067483663559, + 0.029482824727892876, + 0.09246966987848282 + ], + "credit_map": [ + 0.008281506597995758, + 0.009242067113518715, + 0.007710011675953865, + 0.009978431276977062, + 0.009644247591495514, + 0.007638778071850538, + 0.010267163626849651, + 0.011487445794045925, + 0.02216365747153759, + 0.020600860938429832, + 0.008368338458240032, + 0.01570362225174904, + 0.013460492715239525, + 0.006617283448576927, + 0.014062879607081413, + 0.009927588514983654, + 0.01060707587748766, + 0.01704353652894497, + 0.010982229374349117, + 0.011226769536733627, + 0.00847784336656332, + 0.014363499358296394, + 0.010693707503378391, + 0.012781520374119282, + 0.019592322409152985, + 0.021119410172104836, + 0.031345196068286896, + 0.022269438952207565, + 0.017380241304636, + 0.036162007600069046, + 0.023559236899018288, + 0.03162761405110359, + 0.01003227662295103, + 0.01098032109439373, + 0.010216877795755863, + 0.008414232172071934, + 0.006064038723707199, + 0.009892135858535767, + 0.00919877365231514, + 0.005969294812530279, + 0.007066978141665459, + 0.010133487172424793, + 0.01471323799341917, + 0.009538930840790272, + 0.006023409776389599, + 0.007326752878725529, + 0.0083800433203578, + 0.008784942328929901, + 0.014906165190041065, + 0.028464745730161667, + 0.020487505942583084, + 0.019029296934604645, + 0.016856135800480843, + 0.016403239220380783, + 0.020648779347538948, + 0.017309250310063362, + 0.011520544067025185, + 0.010149013251066208, + 0.012648758478462696, + 0.015980662778019905, + 0.014239703305065632, + 0.009253476746380329, + 0.01238540280610323, + 0.010643117129802704, + 0.004297970794141293, + 0.005905803758651018, + 0.0030560288578271866, + 0.0027008443139493465, + 0.0024123764596879482, + 0.00289597874507308, + 0.00363399600610137, + 0.004579827655106783, + 0.006460549309849739, + 0.011625416576862335, + 0.010777567513287067, + 0.013530946336686611, + 0.009235315024852753, + 0.010289636440575123, + 0.014696086756885052, + 0.015854155644774437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.005963864270597696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036214501596987247 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036214501596987247, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036214501596987247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11569339781999588, + 0.08952842652797699, + 0.1566600501537323, + 0.18619433045387268, + 0.08346764743328094, + 0.077630415558815, + 0.11204718053340912, + 0.025557715445756912, + 0.05805236101150513, + 0.09516841173171997 + ], + "credit_map": [ + 0.009807238355278969, + 0.017204448580741882, + 0.013402517884969711, + 0.022262893617153168, + 0.012195641174912453, + 0.01210081297904253, + 0.01400237251073122, + 0.014717470854520798, + 0.007986796088516712, + 0.010233442299067974, + 0.010123716667294502, + 0.013792360201478004, + 0.012912651523947716, + 0.005371118430048227, + 0.02199205756187439, + 0.007116286549717188, + 0.015385409817099571, + 0.012184742838144302, + 0.05139297991991043, + 0.012478427961468697, + 0.008183837868273258, + 0.016652008518576622, + 0.014715428464114666, + 0.025667207315564156, + 0.023689135909080505, + 0.014078167267143726, + 0.03869527205824852, + 0.02509479969739914, + 0.010774350725114346, + 0.028057202696800232, + 0.021182818338274956, + 0.024622583761811256, + 0.012507278472185135, + 0.009211857803165913, + 0.008258819580078125, + 0.020261896774172783, + 0.006547448690980673, + 0.009495614096522331, + 0.010900718159973621, + 0.006284010596573353, + 0.007339378818869591, + 0.013964381068944931, + 0.014905504882335663, + 0.009018823504447937, + 0.005813696887344122, + 0.012558124959468842, + 0.006216343492269516, + 0.007814161479473114, + 0.011901933699846268, + 0.01131529826670885, + 0.013858072459697723, + 0.014471599832177162, + 0.01112738810479641, + 0.014702782966196537, + 0.02222389727830887, + 0.012446211650967598, + 0.0017841022927314043, + 0.0029186620377004147, + 0.0033695106394588947, + 0.00537734804674983, + 0.002881624735891819, + 0.002308605471625924, + 0.0027424499858170748, + 0.0041754115372896194, + 0.00759837357327342, + 0.010232219472527504, + 0.004746141377836466, + 0.005488719791173935, + 0.007490996271371841, + 0.006153804715722799, + 0.007378801237791777, + 0.00896330364048481, + 0.00791629496961832, + 0.016315970569849014, + 0.007841156795620918, + 0.010737205855548382, + 0.013712864369153976, + 0.010105396620929241, + 0.014616033062338829, + 0.01392348762601614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0036214501596987247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13133329153060913, + 0.0750451385974884, + 0.12337396293878555, + 0.05425010249018669, + 0.09599919617176056, + 0.09899181127548218, + 0.12864252924919128, + 0.06235684081912041, + 0.04668519273400307, + 0.18332192301750183 + ], + "credit_map": [ + 0.012933466583490372, + 0.015504531562328339, + 0.017318250611424446, + 0.021024856716394424, + 0.01847933605313301, + 0.012387641705572605, + 0.014363167807459831, + 0.019322028383612633, + 0.00763116218149662, + 0.01058664545416832, + 0.007703911047428846, + 0.0088412556797266, + 0.010581486858427525, + 0.006299405824393034, + 0.013743703253567219, + 0.00965756643563509, + 0.013273952528834343, + 0.015347021631896496, + 0.0226109828799963, + 0.015154375694692135, + 0.009359404444694519, + 0.01680244319140911, + 0.013216059654951096, + 0.017609721049666405, + 0.005228967405855656, + 0.0054391720332205296, + 0.008161818608641624, + 0.005956675391644239, + 0.006394017022103071, + 0.010093177668750286, + 0.006363763473927975, + 0.006612509023398161, + 0.011493622325360775, + 0.016103042289614677, + 0.011325876228511333, + 0.01777028478682041, + 0.0083573954179883, + 0.008985226973891258, + 0.014556385576725006, + 0.007407357916235924, + 0.009713524952530861, + 0.01258421316742897, + 0.013672524131834507, + 0.010705213062465191, + 0.010423801839351654, + 0.012547013349831104, + 0.013790101744234562, + 0.015555422753095627, + 0.013910870999097824, + 0.01605556160211563, + 0.017096709460020065, + 0.014451847411692142, + 0.01216679997742176, + 0.015508671291172504, + 0.02397695556282997, + 0.01547511387616396, + 0.006280985195189714, + 0.008698520250618458, + 0.009931515902280807, + 0.008700764738023281, + 0.005961233749985695, + 0.007503790315240622, + 0.00868725124746561, + 0.006592782214283943, + 0.0057851532474160194, + 0.010161145590245724, + 0.004312438424676657, + 0.003437151201069355, + 0.004239710513502359, + 0.00429318193346262, + 0.0057329717092216015, + 0.008723440580070019, + 0.012110395357012749, + 0.026802122592926025, + 0.019994188100099564, + 0.024877918884158134, + 0.020992731675505638, + 0.017889846116304398, + 0.030133385211229324, + 0.0305213313549757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.020577168092131615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307139292359352 + ], + "opposite_logprob_tokens": [ + -3.1680715084075928 + ], + "logprob_reward": -0.04307139292359352, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307139292359352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14592552185058594, + 0.11557232588529587, + 0.08537687361240387, + 0.13977192342281342, + 0.07905750721693039, + 0.08200410008430481, + 0.15716585516929626, + 0.043874163180589676, + 0.03425351530313492, + 0.11699821054935455 + ], + "credit_map": [ + 0.011572225019335747, + 0.01852223090827465, + 0.01905028708279133, + 0.030165541917085648, + 0.01719171740114689, + 0.013437633402645588, + 0.016402941197156906, + 0.019582942128181458, + 0.019882651045918465, + 0.013382123783230782, + 0.010849399492144585, + 0.0189055185765028, + 0.01595383882522583, + 0.009055081754922867, + 0.016516096889972687, + 0.011027606204152107, + 0.009380333125591278, + 0.011775226332247257, + 0.011777706444263458, + 0.009193727746605873, + 0.007367640733718872, + 0.012196453288197517, + 0.012368245981633663, + 0.011317544616758823, + 0.018131989985704422, + 0.015089835040271282, + 0.02923397906124592, + 0.013638809323310852, + 0.011983992531895638, + 0.02188025414943695, + 0.01567668281495571, + 0.014136383309960365, + 0.014411669224500656, + 0.006780633702874184, + 0.007760224863886833, + 0.02481733076274395, + 0.004564532544463873, + 0.006194727495312691, + 0.009362638927996159, + 0.0051657529547810555, + 0.008260931819677353, + 0.011839086189866066, + 0.00919974222779274, + 0.007837587036192417, + 0.009123419411480427, + 0.011832603253424168, + 0.011876994743943214, + 0.012033736333251, + 0.023822329938411713, + 0.01550229825079441, + 0.033943381160497665, + 0.014941485598683357, + 0.013257051818072796, + 0.01937403902411461, + 0.02037801593542099, + 0.01594724878668785, + 0.0064537543803453445, + 0.006120037753134966, + 0.0058006709441542625, + 0.005339117255061865, + 0.0063322181813418865, + 0.003974945284426212, + 0.0052492450922727585, + 0.004604175686836243, + 0.004962318576872349, + 0.004502498544752598, + 0.00407852977514267, + 0.0029343299102038145, + 0.00444256654009223, + 0.004057341255247593, + 0.004161457996815443, + 0.005114472005516291, + 0.01129896566271782, + 0.020627222955226898, + 0.011020774953067303, + 0.011871549300849438, + 0.013748797588050365, + 0.0135698476806283, + 0.018587958067655563, + 0.016273092478513718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.04307139292359352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, low-angle, wide-angle shot focuses on a downpour-soaked asphalt street in a cyberpunk city. Between towering skyscrapers, giant, translucent holographic advertisements flicker with shifting patterns, interplaying with pink and cyan neon signs, painting the slippery road surface into a shimmering mirror. The air is thick with damp vapor. A heavily modified off-road vehicle suddenly bursts into view, its matte black armored body covered in battle-worn scratches and mud stains. The blue LED strips on its underbody cast an eerie halo of light on the ground. The vehicle drifts with incredible speed and precision. The wide, heavily treaded off-road tires lose grip the moment they make contact, emitting a sharp, grinding sound, and the vehicle tilts violently under the immense force of inertia. At this moment, the high-speed spinning rear wheel suddenly rolled up the accumulated water on the road, exploding a huge fan-shaped water curtain that almost occupied half of the screen. Every splashing water droplet reflected a diamond-like bright and short-lived colorful light under the illumination of the surrounding chaotic light. The whole scene was full of movement and power. Against the backdrop of the cold rainy night and the dazzling technological light pollution, it created a tense, chaotic and visually impactful futuristic atmosphere.", + "seed": 8683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_8403_seed_8683_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026545894145965577, + "vlm_logprob": -0.026545894145965577, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.026545894145965577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017024332657456398 + ], + "opposite_logprob_tokens": [ + -4.267024517059326 + ], + "logprob_reward": -0.017024332657456398, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017024332657456398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31059563159942627, + 0.13004674017429352, + 0.14737030863761902, + 0.0685107484459877, + 0.01660715602338314, + 0.08644048869609833, + 0.03165653347969055, + 0.07946168631315231, + 0.09884418547153473, + 0.030466480180621147 + ], + "credit_map": [ + 0.029089665040373802, + 0.0279565267264843, + 0.04243236407637596, + 0.039395514875650406, + 0.034099847078323364, + 0.046583618968725204, + 0.04247935861349106, + 0.04855876788496971, + 0.008199108764529228, + 0.03284013643860817, + 0.009553671814501286, + 0.016371656209230423, + 0.016866212710738182, + 0.013088085688650608, + 0.019564731046557426, + 0.013563152402639389, + 0.0258259866386652, + 0.008931963704526424, + 0.010589521378278732, + 0.0075147030875086784, + 0.01468050479888916, + 0.016045115888118744, + 0.01457803975790739, + 0.04920447990298271, + 0.009881553240120411, + 0.012665613554418087, + 0.007537350989878178, + 0.006907246075570583, + 0.01069593895226717, + 0.007976428605616093, + 0.006995466072112322, + 0.005851156543940306, + 0.002293979050591588, + 0.0027256435714662075, + 0.002184751909226179, + 0.002549382857978344, + 0.0015762822004035115, + 0.0014276094734668732, + 0.0016261745477095246, + 0.002223334042355418, + 0.014176815748214722, + 0.00273248297162354, + 0.0028208382427692413, + 0.003289474407210946, + 0.0030527650378644466, + 0.036601752042770386, + 0.012136132456362247, + 0.011630230583250523, + 0.0038122220430523157, + 0.0038112702313810587, + 0.00404139282181859, + 0.003102297894656658, + 0.004419087897986174, + 0.003264858154579997, + 0.003996247425675392, + 0.0052091581746935844, + 0.004630046896636486, + 0.010806147940456867, + 0.002999997464939952, + 0.0029370554257184267, + 0.022237282246351242, + 0.006313450168818235, + 0.004088628571480513, + 0.025449085980653763, + 0.03103617951273918, + 0.03585837036371231, + 0.003923028241842985, + 0.004769796971231699, + 0.006976757198572159, + 0.008662758395075798, + 0.0034099705517292023, + 0.004207339137792587, + 0.0025472277775406837, + 0.005002130288630724, + 0.004282355308532715, + 0.0030848339665681124, + 0.0043759336695075035, + 0.003684157272800803, + 0.0034805829636752605, + 0.004009260796010494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.017024332657456398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059593613259494305 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.0059593613259494305, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059593613259494305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05847422033548355, + 0.06332019716501236, + 0.08228641748428345, + 0.039314933121204376, + 0.035893164575099945, + 0.22184304893016815, + 0.04232219606637955, + 0.30423128604888916, + 0.04258563369512558, + 0.1097288727760315 + ], + "credit_map": [ + 0.004628047347068787, + 0.009237153455615044, + 0.008103358559310436, + 0.005580865312367678, + 0.009767135605216026, + 0.005668873433023691, + 0.007728802040219307, + 0.007759986910969019, + 0.004204371012747288, + 0.014676572754979134, + 0.00525969173759222, + 0.009267858229577541, + 0.009378863498568535, + 0.006226476281881332, + 0.008617956191301346, + 0.005688404198735952, + 0.008758044801652431, + 0.005516259465366602, + 0.006425369996577501, + 0.006962309591472149, + 0.01997797191143036, + 0.006780853495001793, + 0.00889834575355053, + 0.018967261537909508, + 0.004519945476204157, + 0.006033715792000294, + 0.00398866506293416, + 0.005402394570410252, + 0.005839975085109472, + 0.0046242703683674335, + 0.005330659449100494, + 0.0035753061529248953, + 0.003881504526361823, + 0.0038183056749403477, + 0.004018912557512522, + 0.005491769406944513, + 0.004246068187057972, + 0.003966459073126316, + 0.004956186283379793, + 0.005513956304639578, + 0.031930312514305115, + 0.00692872004583478, + 0.006765926256775856, + 0.009891184978187084, + 0.008636188693344593, + 0.09311796724796295, + 0.030991114675998688, + 0.03358162194490433, + 0.004844340030103922, + 0.006594894919544458, + 0.004189544823020697, + 0.004611039999872446, + 0.005855504889041185, + 0.004117561038583517, + 0.004924533423036337, + 0.0071847750805318356, + 0.015663452446460724, + 0.03928535059094429, + 0.011724726296961308, + 0.01650172472000122, + 0.08212628215551376, + 0.02418423257768154, + 0.02201765961945057, + 0.09272786229848862, + 0.014951550401747227, + 0.012637528590857983, + 0.0020405123941600323, + 0.0015958509175106883, + 0.003458762774243951, + 0.0035135536454617977, + 0.002294955775141716, + 0.0020929169841110706, + 0.010086191818118095, + 0.01827879436314106, + 0.01582270674407482, + 0.011403138749301434, + 0.017121901735663414, + 0.011277133598923683, + 0.010401480831205845, + 0.015337524004280567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.0059593613259494305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639812279492617 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639812279492617, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639812279492617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048332955688238144, + 0.09445814043283463, + 0.040423545986413956, + 0.030421659350395203, + 0.018358001485466957, + 0.18809404969215393, + 0.02893722988665104, + 0.2842715382575989, + 0.22291769087314606, + 0.04378523677587509 + ], + "credit_map": [ + 0.007124345283955336, + 0.008562830276787281, + 0.0065963477827608585, + 0.0038506616838276386, + 0.006265577394515276, + 0.00547814229503274, + 0.005194122437387705, + 0.00526092667132616, + 0.005916359834372997, + 0.02091757394373417, + 0.00699732918292284, + 0.010971794836223125, + 0.012306773103773594, + 0.009119226597249508, + 0.018045956268906593, + 0.010183123871684074, + 0.0058677406050264835, + 0.0038803196512162685, + 0.004299528896808624, + 0.004430797416716814, + 0.005017283372581005, + 0.0036544809117913246, + 0.003414275124669075, + 0.009859117679297924, + 0.00454731797799468, + 0.0043764724396169186, + 0.0032984307035803795, + 0.0050085317343473434, + 0.003267008578404784, + 0.0035385240335017443, + 0.004061135463416576, + 0.002324239118024707, + 0.002261267276480794, + 0.0031175464391708374, + 0.0027052764780819416, + 0.0030696173198521137, + 0.0015641333302482963, + 0.001628064550459385, + 0.0013466734671965241, + 0.002665423322468996, + 0.028502751141786575, + 0.005614513531327248, + 0.006367295049130917, + 0.007798716891556978, + 0.005720961838960648, + 0.08074438571929932, + 0.020303038880228996, + 0.03304240107536316, + 0.0036033443175256252, + 0.0077177975326776505, + 0.0017007114365696907, + 0.0032927398569881916, + 0.0038883432280272245, + 0.0027224088553339243, + 0.0029750631656497717, + 0.0030368200968950987, + 0.015264179557561874, + 0.03808417171239853, + 0.008677484467625618, + 0.009610945358872414, + 0.08224699646234512, + 0.022161060944199562, + 0.013541033491492271, + 0.09468565136194229, + 0.087958924472332, + 0.07496190071105957, + 0.008757606148719788, + 0.009085600264370441, + 0.01374171394854784, + 0.012426819652318954, + 0.007939266040921211, + 0.008045854046940804, + 0.003808249020949006, + 0.00678064301609993, + 0.005293837748467922, + 0.006365669891238213, + 0.005013501271605492, + 0.004997530020773411, + 0.006199854426085949, + 0.005325951613485813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.004639812279492617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606660895049572 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.01606660895049572, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606660895049572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15578033030033112, + 0.034465618431568146, + 0.06991156935691833, + 0.04069396108388901, + 0.027533741667866707, + 0.029719067737460136, + 0.09056053310632706, + 0.4362860321998596, + 0.07286687940359116, + 0.04218224808573723 + ], + "credit_map": [ + 0.012641927227377892, + 0.01934349723160267, + 0.010630512610077858, + 0.015977920964360237, + 0.011715704575181007, + 0.034541793167591095, + 0.019666284322738647, + 0.03126269578933716, + 0.0029233191162347794, + 0.007545894477516413, + 0.0029190238565206528, + 0.004695775918662548, + 0.004340223968029022, + 0.0032686390914022923, + 0.005351293832063675, + 0.003421447705477476, + 0.006834239233285189, + 0.005806913133710623, + 0.008120004087686539, + 0.008680437691509724, + 0.018814900889992714, + 0.0055718873627483845, + 0.009314245544373989, + 0.0067689428105950356, + 0.0055809421464800835, + 0.007866945117712021, + 0.0044769709929823875, + 0.005814956501126289, + 0.0049841152504086494, + 0.00320484209805727, + 0.005183512810617685, + 0.0035816754680126905, + 0.0033152990508824587, + 0.0031585879623889923, + 0.003461886662989855, + 0.004513939376920462, + 0.002634271513670683, + 0.0026231103111058474, + 0.0022896388545632362, + 0.005537006072700024, + 0.002835844876244664, + 0.00359308416955173, + 0.0029721539467573166, + 0.0028414686676114798, + 0.0030940219294279814, + 0.004042025655508041, + 0.00494280643761158, + 0.00539766252040863, + 0.011342794634401798, + 0.015309629961848259, + 0.00606641685590148, + 0.011880640871822834, + 0.01831842213869095, + 0.008219410665333271, + 0.010324562899768353, + 0.00909865740686655, + 0.03948473930358887, + 0.05138106644153595, + 0.023082586005330086, + 0.01865757629275322, + 0.11361869424581528, + 0.03230922669172287, + 0.02685084380209446, + 0.1309012919664383, + 0.017811693251132965, + 0.005273632705211639, + 0.0042881169356405735, + 0.0035317970905452967, + 0.004327473696321249, + 0.028393184766173363, + 0.004306266084313393, + 0.004934715107083321, + 0.003895876929163933, + 0.008261959999799728, + 0.005548406392335892, + 0.0048575433902442455, + 0.004469939973205328, + 0.004673340357840061, + 0.003920451737940311, + 0.006554727908223867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01606660895049572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903935551643372 + ], + "opposite_logprob_tokens": [ + -2.4640393257141113 + ], + "logprob_reward": -0.08903935551643372, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903935551643372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05738960951566696, + 0.055203765630722046, + 0.058501821011304855, + 0.029843905940651894, + 0.020594991743564606, + 0.23982444405555725, + 0.025737233459949493, + 0.2218090295791626, + 0.2613106966018677, + 0.02978449873626232 + ], + "credit_map": [ + 0.004354231525212526, + 0.0077882553450763226, + 0.004831038881093264, + 0.008104079402983189, + 0.006466371938586235, + 0.0075724865309894085, + 0.009821279905736446, + 0.008451864123344421, + 0.0033406938891857862, + 0.010847058147192001, + 0.004434104543179274, + 0.008335467427968979, + 0.007877888157963753, + 0.004776518326252699, + 0.009027063846588135, + 0.006564972922205925, + 0.007444437593221664, + 0.006321433000266552, + 0.004308978095650673, + 0.004586776718497276, + 0.019195040687918663, + 0.005025675520300865, + 0.007555633783340454, + 0.004063842352479696, + 0.0038954003248363733, + 0.004076378885656595, + 0.0025971350260078907, + 0.0045608459040522575, + 0.003558172844350338, + 0.002998682903125882, + 0.004533054772764444, + 0.0036242378409951925, + 0.002537520369514823, + 0.003515387885272503, + 0.0024889588821679354, + 0.0020087058655917645, + 0.002233626786619425, + 0.003347277408465743, + 0.0018466741312295198, + 0.002616840647533536, + 0.04027056694030762, + 0.007464912254363298, + 0.008736157789826393, + 0.007600066252052784, + 0.008129353635013103, + 0.10270076245069504, + 0.028184587135910988, + 0.03673803433775902, + 0.002514574211090803, + 0.005445157177746296, + 0.0016464180080220103, + 0.003308243816718459, + 0.003658085595816374, + 0.0021683787927031517, + 0.0038678909186273813, + 0.003128485521301627, + 0.01111620757728815, + 0.030251547694206238, + 0.006805041339248419, + 0.007398412097245455, + 0.06438412517309189, + 0.01760534755885601, + 0.010258717462420464, + 0.0739896297454834, + 0.10375560075044632, + 0.0857081189751625, + 0.010042496956884861, + 0.009316056966781616, + 0.01655622012913227, + 0.016890710219740868, + 0.010007956996560097, + 0.00903352815657854, + 0.0027348450385034084, + 0.0052484143525362015, + 0.0038455629255622625, + 0.003817725693807006, + 0.0033212986309081316, + 0.0034505128860473633, + 0.0029193791560828686, + 0.004446759819984436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.08903935551643372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, low-angle, wide-angle shot focuses on a downpour-soaked asphalt street in a cyberpunk city. Between towering skyscrapers, giant, translucent holographic advertisements flicker with shifting patterns, interplaying with pink and cyan neon signs, painting the slippery road surface into a shimmering mirror. The air is thick with damp vapor. A heavily modified off-road vehicle suddenly bursts into view, its matte black armored body covered in battle-worn scratches and mud stains. The blue LED strips on its underbody cast an eerie halo of light on the ground. The vehicle drifts with incredible speed and precision. The wide, heavily treaded off-road tires lose grip the moment they make contact, emitting a sharp, grinding sound, and the vehicle tilts violently under the immense force of inertia. At this moment, the high-speed spinning rear wheel suddenly rolled up the accumulated water on the road, exploding a huge fan-shaped water curtain that almost occupied half of the screen. Every splashing water droplet reflected a diamond-like bright and short-lived colorful light under the illumination of the surrounding chaotic light. The whole scene was full of movement and power. Against the backdrop of the cold rainy night and the dazzling technological light pollution, it created a tense, chaotic and visually impactful futuristic atmosphere.", + "seed": 8684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_8403_seed_8684_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018677817191928625, + "vlm_logprob": -0.018677817191928625, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.018677817191928625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00935598649084568 + ], + "opposite_logprob_tokens": [ + -5.2593560218811035 + ], + "logprob_reward": -0.00935598649084568, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00935598649084568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16145645081996918, + 0.07500100880861282, + 0.04476114362478256, + 0.0719594731926918, + 0.10171873867511749, + 0.027623986825346947, + 0.12134210765361786, + 0.1786782443523407, + 0.05382835492491722, + 0.16363045573234558 + ], + "credit_map": [ + 0.016815444454550743, + 0.02285122498869896, + 0.02558378130197525, + 0.02552425116300583, + 0.01494793500751257, + 0.017980288714170456, + 0.02130669169127941, + 0.016446813941001892, + 0.012749291956424713, + 0.012048127129673958, + 0.015482673421502113, + 0.011362267658114433, + 0.006394348107278347, + 0.0056045763194561005, + 0.0063881720416247845, + 0.004971555434167385, + 0.004606351722031832, + 0.005386997014284134, + 0.008446422405540943, + 0.006588696036487818, + 0.008038956671953201, + 0.005007200408726931, + 0.003725420217961073, + 0.0029610982164740562, + 0.006242738105356693, + 0.006234962958842516, + 0.00829156395047903, + 0.007777341175824404, + 0.013005069456994534, + 0.010573919862508774, + 0.00986380036920309, + 0.00997007917612791, + 0.010035104118287563, + 0.016139773651957512, + 0.011415313929319382, + 0.013967872597277164, + 0.015130329877138138, + 0.009194700978696346, + 0.0143293347209692, + 0.01150631532073021, + 0.0036368865985423326, + 0.004245465155690908, + 0.003182771848514676, + 0.002471508225426078, + 0.0028657468501478434, + 0.0027504214085638523, + 0.0028483939822763205, + 0.005622790660709143, + 0.013412910513579845, + 0.013983258977532387, + 0.015566006302833557, + 0.02267678827047348, + 0.017896929755806923, + 0.011317052878439426, + 0.014915578067302704, + 0.011573582887649536, + 0.01492683682590723, + 0.049354005604982376, + 0.059758830815553665, + 0.024010371416807175, + 0.008226761594414711, + 0.007065068930387497, + 0.00866707507520914, + 0.006669300142675638, + 0.009582306258380413, + 0.006672271527349949, + 0.0076398118399083614, + 0.008852662518620491, + 0.004239077214151621, + 0.004218808375298977, + 0.007999378256499767, + 0.0046240417286753654, + 0.008511006832122803, + 0.009547919034957886, + 0.035753149539232254, + 0.052946582436561584, + 0.012919684872031212, + 0.010483704507350922, + 0.01853567734360695, + 0.014932718127965927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00935598649084568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548067606985569 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012548067606985569, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548067606985569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15697181224822998, + 0.16798236966133118, + 0.0623018741607666, + 0.0954134464263916, + 0.07394976913928986, + 0.12128852307796478, + 0.11924052238464355, + 0.11408919841051102, + 0.055539779365062714, + 0.03322269767522812 + ], + "credit_map": [ + 0.017553703859448433, + 0.019924387335777283, + 0.022002262994647026, + 0.024196496233344078, + 0.02058282308280468, + 0.01292668841779232, + 0.024201741442084312, + 0.015583711676299572, + 0.015846529975533485, + 0.0315869115293026, + 0.02085687220096588, + 0.03555800020694733, + 0.016896046698093414, + 0.01611454039812088, + 0.01745361089706421, + 0.013669852167367935, + 0.007057668641209602, + 0.0082772271707654, + 0.011627034284174442, + 0.009923810139298439, + 0.005267294589430094, + 0.008380942977964878, + 0.005994664039462805, + 0.005773234646767378, + 0.007212023250758648, + 0.008869733661413193, + 0.011246507056057453, + 0.011161932721734047, + 0.018731044605374336, + 0.013772276230156422, + 0.012689540162682533, + 0.011730390600860119, + 0.008425846695899963, + 0.008637549355626106, + 0.011847754940390587, + 0.01091602724045515, + 0.007794930599629879, + 0.007272223010659218, + 0.00969287846237421, + 0.009362563490867615, + 0.019353695213794708, + 0.016195544973015785, + 0.013190113008022308, + 0.015252822078764439, + 0.014483681879937649, + 0.012778897769749165, + 0.015981154516339302, + 0.0140526182949543, + 0.008853795006871223, + 0.011647440493106842, + 0.01459077000617981, + 0.029383433982729912, + 0.016576798632740974, + 0.00963007751852274, + 0.014056372456252575, + 0.014501838944852352, + 0.008810735307633877, + 0.02753056399524212, + 0.03461018204689026, + 0.01678759790956974, + 0.006762052420526743, + 0.006995563395321369, + 0.006748448126018047, + 0.005844052415341139, + 0.003661162219941616, + 0.005326366517692804, + 0.012708982452750206, + 0.009677491150796413, + 0.003447070484980941, + 0.0036240527406334877, + 0.010480028577148914, + 0.0066146221943199635, + 0.0026248465292155743, + 0.003042735392227769, + 0.008699045516550541, + 0.00754312239587307, + 0.0017655862029641867, + 0.0029147490859031677, + 0.0036954462993890047, + 0.0029371646232903004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.012548067606985569 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17713266611099243, + 0.14430031180381775, + 0.10846293717622757, + 0.10651523619890213, + 0.14888283610343933, + 0.03246693313121796, + 0.02806650660932064, + 0.10301940888166428, + 0.06796226650476456, + 0.0831909030675888 + ], + "credit_map": [ + 0.013573120348155499, + 0.021760931238532066, + 0.022729650139808655, + 0.03539220616221428, + 0.012831195257604122, + 0.015619209036231041, + 0.03536711633205414, + 0.019859246909618378, + 0.017586948350071907, + 0.025780195370316505, + 0.02211068756878376, + 0.02511918544769287, + 0.013283823616802692, + 0.01194748468697071, + 0.016886690631508827, + 0.011585297994315624, + 0.010043288581073284, + 0.010079636238515377, + 0.020193375647068024, + 0.01735123060643673, + 0.011570148169994354, + 0.013408128172159195, + 0.01655603013932705, + 0.009261103346943855, + 0.009276598691940308, + 0.008204573765397072, + 0.013167661614716053, + 0.013484320603311062, + 0.017471760511398315, + 0.014669962227344513, + 0.016246240586042404, + 0.013994120061397552, + 0.014838901348412037, + 0.020371800288558006, + 0.017771277576684952, + 0.021588213741779327, + 0.022985555231571198, + 0.014081013388931751, + 0.021906815469264984, + 0.015339257195591927, + 0.005031903274357319, + 0.0031078984029591084, + 0.003421199508011341, + 0.003263064194470644, + 0.0033608556259423494, + 0.00346794119104743, + 0.0035503297112882137, + 0.00726374052464962, + 0.0031786179170012474, + 0.0035746791400015354, + 0.0025554781313985586, + 0.004799956921488047, + 0.004026349633932114, + 0.00255178683437407, + 0.004689257126301527, + 0.002690379973500967, + 0.008352098055183887, + 0.028999725356698036, + 0.03546081483364105, + 0.013416492380201817, + 0.005105067044496536, + 0.003950850572437048, + 0.004044061992317438, + 0.003690296784043312, + 0.003309621475636959, + 0.004298523534089327, + 0.024299610406160355, + 0.0204111710190773, + 0.002783628646284342, + 0.00277466862462461, + 0.00593628128990531, + 0.004148765001446009, + 0.003746338188648224, + 0.003199170110747218, + 0.029189396649599075, + 0.03145812824368477, + 0.0032780354376882315, + 0.003193790791556239, + 0.005491639021784067, + 0.0036344078835099936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.00464016804471612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181568011641502 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181568011641502, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181568011641502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09730523824691772, + 0.047533903270959854, + 0.05844062939286232, + 0.0397125668823719, + 0.049877915531396866, + 0.025429602712392807, + 0.38055920600891113, + 0.03867894783616066, + 0.04371102899312973, + 0.2187509685754776 + ], + "credit_map": [ + 0.007603469304740429, + 0.012662980705499649, + 0.010549575090408325, + 0.01785460114479065, + 0.007266636937856674, + 0.009826148860156536, + 0.020128857344388962, + 0.0114129688590765, + 0.005676845088601112, + 0.007533888332545757, + 0.008917761966586113, + 0.007581001613289118, + 0.004068965092301369, + 0.0036144512705504894, + 0.0060666645877063274, + 0.004074322059750557, + 0.006524119526147842, + 0.008269826881587505, + 0.009750197641551495, + 0.007090530823916197, + 0.011276336386799812, + 0.005718324799090624, + 0.004959731828421354, + 0.004851565696299076, + 0.003025586949661374, + 0.002746701706200838, + 0.003750292584300041, + 0.004674926865845919, + 0.006186436861753464, + 0.005530539434403181, + 0.0074288914911448956, + 0.0063691893592476845, + 0.004515953361988068, + 0.007407146040350199, + 0.005462097004055977, + 0.004829699173569679, + 0.006665688008069992, + 0.004444365855306387, + 0.007093383464962244, + 0.009459581226110458, + 0.005320414435118437, + 0.003019683761522174, + 0.003156929509714246, + 0.002466750331223011, + 0.0027330394368618727, + 0.00269287615083158, + 0.0026061078533530235, + 0.0034338023979216814, + 0.06422031670808792, + 0.047410160303115845, + 0.04472128674387932, + 0.07773648947477341, + 0.038854822516441345, + 0.04039425030350685, + 0.0404767245054245, + 0.026745161041617393, + 0.008042980916798115, + 0.0054113301448524, + 0.004167987033724785, + 0.005984682124108076, + 0.002955797128379345, + 0.004660166334360838, + 0.003653664840385318, + 0.003802339779213071, + 0.0027667758986353874, + 0.00364780449308455, + 0.011414515785872936, + 0.008437009528279305, + 0.0026820471975952387, + 0.0030596130527555943, + 0.006648397073149681, + 0.005054865498095751, + 0.008555997163057327, + 0.0055937375873327255, + 0.07387720793485641, + 0.09093356132507324, + 0.00519597576931119, + 0.004559538327157497, + 0.0174369215965271, + 0.012598028406500816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.018181568011641502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663295805454254 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.048663295805454254, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663295805454254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13418391346931458, + 0.07155479490756989, + 0.04891221225261688, + 0.05415606498718262, + 0.09477004408836365, + 0.030270390212535858, + 0.080940380692482, + 0.30276399850845337, + 0.07217631489038467, + 0.11027190834283829 + ], + "credit_map": [ + 0.009678398258984089, + 0.01858813688158989, + 0.014054004102945328, + 0.02608494460582733, + 0.010485579259693623, + 0.013439578004181385, + 0.02488652989268303, + 0.016966737806797028, + 0.008359109982848167, + 0.014011722058057785, + 0.011378565803170204, + 0.013017542660236359, + 0.005483612418174744, + 0.005865569692105055, + 0.007709025405347347, + 0.0057296473532915115, + 0.005940646398812532, + 0.006212308071553707, + 0.009094448760151863, + 0.007274797651916742, + 0.005211857613176107, + 0.005524228326976299, + 0.005826862994581461, + 0.003827060107141733, + 0.004758915863931179, + 0.0036402365658432245, + 0.00783159676939249, + 0.0072122481651604176, + 0.00781894288957119, + 0.0070860073901712894, + 0.008326656185090542, + 0.007481455337256193, + 0.008333257399499416, + 0.011756784282624722, + 0.010734623298048973, + 0.011533898301422596, + 0.014094945043325424, + 0.008699523285031319, + 0.014747139066457748, + 0.01486987341195345, + 0.0045128874480724335, + 0.0038412923458963633, + 0.003555855480954051, + 0.0029815665911883116, + 0.003647406818345189, + 0.0033774308394640684, + 0.0031671274919062853, + 0.005186822731047869, + 0.00747451838105917, + 0.00918706227093935, + 0.010439164936542511, + 0.01957329921424389, + 0.008416163735091686, + 0.007140816655009985, + 0.01024293340742588, + 0.008466411381959915, + 0.023569339886307716, + 0.08964453637599945, + 0.1062937006354332, + 0.036558546125888824, + 0.013763455674052238, + 0.011606973595917225, + 0.010302746668457985, + 0.011024688370525837, + 0.006024912931025028, + 0.008511219173669815, + 0.017995087429881096, + 0.01519132498651743, + 0.005979946348816156, + 0.005902593489736319, + 0.007193617522716522, + 0.005377606023102999, + 0.0054891579784452915, + 0.006467985454946756, + 0.03584258630871773, + 0.039395999163389206, + 0.005933587439358234, + 0.004572358448058367, + 0.007531516719609499, + 0.005038709845393896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.048663295805454254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, low-angle, wide-angle shot focuses on a downpour-soaked asphalt street in a cyberpunk city. Between towering skyscrapers, giant, translucent holographic advertisements flicker with shifting patterns, interplaying with pink and cyan neon signs, painting the slippery road surface into a shimmering mirror. The air is thick with damp vapor. A heavily modified off-road vehicle suddenly bursts into view, its matte black armored body covered in battle-worn scratches and mud stains. The blue LED strips on its underbody cast an eerie halo of light on the ground. The vehicle drifts with incredible speed and precision. The wide, heavily treaded off-road tires lose grip the moment they make contact, emitting a sharp, grinding sound, and the vehicle tilts violently under the immense force of inertia. At this moment, the high-speed spinning rear wheel suddenly rolled up the accumulated water on the road, exploding a huge fan-shaped water curtain that almost occupied half of the screen. Every splashing water droplet reflected a diamond-like bright and short-lived colorful light under the illumination of the surrounding chaotic light. The whole scene was full of movement and power. Against the backdrop of the cold rainy night and the dazzling technological light pollution, it created a tense, chaotic and visually impactful futuristic atmosphere.", + "seed": 8685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_4/index_8403_seed_8685_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027984621096402408, + "vlm_logprob": -0.027984621096402408, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.027984621096402408, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007294686511158943 + ], + "opposite_logprob_tokens": [ + -5.757294654846191 + ], + "logprob_reward": -0.007294686511158943, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007294686511158943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08132150769233704, + 0.06991913914680481, + 0.06312834471464157, + 0.059907544404268265, + 0.2596152424812317, + 0.14152315258979797, + 0.1486198604106903, + 0.07555869221687317, + 0.05140793323516846, + 0.048998553305864334 + ], + "credit_map": [ + 0.011606650426983833, + 0.011819861829280853, + 0.014006233774125576, + 0.009202598594129086, + 0.007036327850073576, + 0.008976664394140244, + 0.009214474819600582, + 0.009458698332309723, + 0.01219511590898037, + 0.0075447577983140945, + 0.006544393487274647, + 0.0072199320420622826, + 0.008160778321325779, + 0.011754545383155346, + 0.009738859720528126, + 0.006760755088180304, + 0.008134630508720875, + 0.009630321525037289, + 0.00753621244803071, + 0.01035875640809536, + 0.0064928168430924416, + 0.006272012367844582, + 0.009580752812325954, + 0.0051228380762040615, + 0.004801122006028891, + 0.009007212705910206, + 0.006692135706543922, + 0.009802008047699928, + 0.008660043589770794, + 0.006748025771230459, + 0.009544888511300087, + 0.004652111791074276, + 0.012145984917879105, + 0.007045550737529993, + 0.020330581814050674, + 0.009189924225211143, + 0.008923803456127644, + 0.09381014108657837, + 0.011671768501400948, + 0.09649746119976044, + 0.01052169594913721, + 0.007708868011832237, + 0.0169537290930748, + 0.007886640727519989, + 0.011641174554824829, + 0.0407404899597168, + 0.007500828243792057, + 0.038569726049900055, + 0.0077858553268015385, + 0.008991356939077377, + 0.010161272250115871, + 0.005633206106722355, + 0.005423746537417173, + 0.054660096764564514, + 0.005523104220628738, + 0.05044122040271759, + 0.007500775624066591, + 0.008044480346143246, + 0.01707969419658184, + 0.008975683711469173, + 0.007682841271162033, + 0.008084076456725597, + 0.005958462134003639, + 0.012232683598995209, + 0.0061211599968373775, + 0.007891621440649033, + 0.007696185261011124, + 0.00586304347962141, + 0.006611311342567205, + 0.009094590321183205, + 0.0047048586420714855, + 0.0034251660108566284, + 0.005498882383108139, + 0.009128964506089687, + 0.006486700382083654, + 0.005485113710165024, + 0.005904616788029671, + 0.008353368379175663, + 0.00362658710218966, + 0.0045143188908696175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.007294686511158943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644795786589384 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004644795786589384, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644795786589384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.003432770259678364, + 0.0032481481321156025, + 0.0034311707131564617, + 0.0024312736932188272, + 0.008889401331543922, + 0.009220114909112453, + 0.9616696238517761, + 0.0017225369811058044, + 0.002519714180380106, + 0.003435251535847783 + ], + "credit_map": [ + 0.00043870252557098866, + 0.0005652238614857197, + 0.0006689949077554047, + 0.00044277298729866743, + 0.00027329957811161876, + 0.0003656090993899852, + 0.00036101514706388116, + 0.00031715162913315, + 0.00037154104211367667, + 0.0003869913052767515, + 0.0002845086273737252, + 0.0003465131449047476, + 0.0003760237595997751, + 0.0006030239746905863, + 0.0005861044046469033, + 0.00029344143695198, + 0.00027876446256414056, + 0.000651136739179492, + 0.0004619637329597026, + 0.0006199989002197981, + 0.0003390365745872259, + 0.00034419717849232256, + 0.00044155167415738106, + 0.00029452104354277253, + 0.0004043139051645994, + 0.00026371271815150976, + 0.0002735232119448483, + 0.0002466462610755116, + 0.0003296154027339071, + 0.0003842902951873839, + 0.0003310672764200717, + 0.00019810420053545386, + 0.0008360289502888918, + 0.0009894970571622252, + 0.0012123642954975367, + 0.0007033499423414469, + 0.0007505340618081391, + 0.001774212927557528, + 0.0005491672200150788, + 0.002074246061965823, + 0.000572436023503542, + 0.000720408046618104, + 0.0011202305322512984, + 0.000606955902185291, + 0.0007994716288521886, + 0.002449152758345008, + 0.0005526655586436391, + 0.0023987931199371815, + 0.03706613555550575, + 0.0361349917948246, + 0.03221369534730911, + 0.022462643682956696, + 0.01847047545015812, + 0.39823752641677856, + 0.027913279831409454, + 0.3891708552837372, + 0.00016624580894131213, + 0.000233707221923396, + 0.0002571707882452756, + 0.00019280266133137047, + 0.0002576415427029133, + 0.00031584760290570557, + 0.00012113399861846119, + 0.00017798709450289607, + 0.0002902199630625546, + 0.0005034304340369999, + 0.0006097842124290764, + 0.00028755926177836955, + 0.0002525602176319808, + 0.00018353169434703887, + 0.0002033118944382295, + 0.00018931616796180606, + 0.00029370022821240127, + 0.0006549185491167009, + 0.00046091648982837796, + 0.00039593648398295045, + 0.0007402832852676511, + 0.0003207148693036288, + 0.0003034632245544344, + 0.00026531788171269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.004644795786589384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252727307379246 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252727307379246, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252727307379246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05476754158735275, + 0.056165020912885666, + 0.045844655483961105, + 0.03301035240292549, + 0.06499355286359787, + 0.054974399507045746, + 0.6007703542709351, + 0.05453948676586151, + 0.01836521551012993, + 0.016569413244724274 + ], + "credit_map": [ + 0.0064818658865988255, + 0.006227234378457069, + 0.009402544237673283, + 0.00824673566967249, + 0.005090564023703337, + 0.004634551703929901, + 0.007828556932508945, + 0.00685548922047019, + 0.005653695203363895, + 0.0056869774125516415, + 0.006300259847193956, + 0.009516684338450432, + 0.007112496066838503, + 0.005658040754497051, + 0.011236335150897503, + 0.0050005316734313965, + 0.004275767132639885, + 0.008996516466140747, + 0.005916668102145195, + 0.008934307843446732, + 0.005557325202971697, + 0.0040643783286213875, + 0.004570674151182175, + 0.0035290191881358624, + 0.003774187294766307, + 0.003543847007676959, + 0.0038994441274553537, + 0.0040178848430514336, + 0.006495104171335697, + 0.004282302223145962, + 0.004742768593132496, + 0.0022548134438693523, + 0.004488819278776646, + 0.0030809547752141953, + 0.006969084497541189, + 0.0035310927778482437, + 0.0026840572245419025, + 0.018243281170725822, + 0.004100014455616474, + 0.021896248683333397, + 0.0050995852798223495, + 0.006101776845753193, + 0.01326061226427555, + 0.005086855497211218, + 0.0074743651784956455, + 0.006218508817255497, + 0.005197080783545971, + 0.006535612046718597, + 0.02319784089922905, + 0.02297142706811428, + 0.020128890872001648, + 0.01462891697883606, + 0.012303070165216923, + 0.24773786962032318, + 0.017191994935274124, + 0.242610365152359, + 0.006372204516083002, + 0.006183944642543793, + 0.007044672966003418, + 0.006055156700313091, + 0.005474532954394817, + 0.015846800059080124, + 0.00333535042591393, + 0.004226823337376118, + 0.0024969587102532387, + 0.003439820371568203, + 0.002189145190641284, + 0.002352384617552161, + 0.0025176533963531256, + 0.0019718047697097063, + 0.001824831240810454, + 0.0015726176789030433, + 0.00158127979375422, + 0.0026252269744873047, + 0.001967640360817313, + 0.0014972459757700562, + 0.0029359476175159216, + 0.0024788284208625555, + 0.0018196922028437257, + 0.0016635521315038204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.005252727307379246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368772566318512 + ], + "opposite_logprob_tokens": [ + -3.4086878299713135 + ], + "logprob_reward": -0.03368772566318512, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368772566318512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014659379608929157, + 0.010437250137329102, + 0.010004867799580097, + 0.007181196939200163, + 0.08084602653980255, + 0.03342930227518082, + 0.8019576072692871, + 0.013175223022699356, + 0.009614282287657261, + 0.018694931641221046 + ], + "credit_map": [ + 0.0014059337554499507, + 0.0020849439315497875, + 0.0027681072242558002, + 0.0022362670861184597, + 0.0012458981946110725, + 0.001338432659395039, + 0.002165701473131776, + 0.0014140954008325934, + 0.0012301703682169318, + 0.0016043653013184667, + 0.0007525811670348048, + 0.0012210211716592312, + 0.0011416703928261995, + 0.0017575477249920368, + 0.0013290653005242348, + 0.0014008274301886559, + 0.0013529674615710974, + 0.0011869166046380997, + 0.00113813451025635, + 0.0010658514220267534, + 0.0010612135520204902, + 0.0012264621909707785, + 0.002001032466068864, + 0.0009722886607050896, + 0.0005645552300848067, + 0.0009151780977845192, + 0.0008018203661777079, + 0.0011909723980352283, + 0.001059742528013885, + 0.0010337529238313437, + 0.0010436199372634292, + 0.0005715549923479557, + 0.0024605232756584883, + 0.0012240662472322583, + 0.002796397777274251, + 0.001838712953031063, + 0.0014879540540277958, + 0.03256653994321823, + 0.0029291354585438967, + 0.03554269298911095, + 0.0025442938786000013, + 0.0017350175185129046, + 0.004575297702103853, + 0.0018863548757508397, + 0.0027321854140609503, + 0.009226896800100803, + 0.00185175403021276, + 0.008877500891685486, + 0.030908240005373955, + 0.03029588982462883, + 0.027241071686148643, + 0.018807047978043556, + 0.015514151193201542, + 0.3319227695465088, + 0.023216096684336662, + 0.32405227422714233, + 0.00148951995652169, + 0.0016064549563452601, + 0.0028682637494057417, + 0.0018082334427163005, + 0.0013337001437321305, + 0.0013537837658077478, + 0.0007863751379773021, + 0.0019288909388706088, + 0.001072934246622026, + 0.001475589582696557, + 0.0019042746862396598, + 0.0011668148217722774, + 0.0009485903428867459, + 0.0011639291187748313, + 0.0008658174192532897, + 0.0010163310216739774, + 0.0023292216937988997, + 0.004882460460066795, + 0.0029912437312304974, + 0.0014961457345634699, + 0.002656888449564576, + 0.0018675450701266527, + 0.0010684613371267915, + 0.0014029624871909618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.03368772566318512 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a downpour-soaked asphalt street in a cyberpunk city with towering skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavily modified off-road vehicle with a matte black armored body covered in scratches and mud visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do blue LED strips on the vehicle's underbody cast a light halo on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spinning rear wheel splash a huge fan-shaped water curtain that occupies approximately half of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot fixed, low-angle, and wide-angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904317021369934 + ], + "opposite_logprob_tokens": [ + -2.464043140411377 + ], + "logprob_reward": -0.08904317021369934, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904317021369934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0074539766646921635, + 0.005781198386102915, + 0.0060799396596848965, + 0.007019594311714172, + 0.019985003396868706, + 0.042516034096479416, + 0.8756492137908936, + 0.016390839591622353, + 0.009988106787204742, + 0.009136077016592026 + ], + "credit_map": [ + 0.00086793745867908, + 0.0009541581384837627, + 0.0017701731994748116, + 0.0010746691841632128, + 0.0005774936289526522, + 0.000683262653183192, + 0.0008312806603498757, + 0.0006950022070668638, + 0.0010415990836918354, + 0.000874014338478446, + 0.0006404320010915399, + 0.0007992462487891316, + 0.0006219713832251728, + 0.0006290213786996901, + 0.0005919731920585036, + 0.0005829411093145609, + 0.000532218546140939, + 0.0009369093459099531, + 0.0007433992577716708, + 0.0008750705164857209, + 0.0007428661338053644, + 0.0007252683863043785, + 0.0008093929500319064, + 0.0007148149306885898, + 0.001149562420323491, + 0.0009179125772789121, + 0.0007765449699945748, + 0.0009604984079487622, + 0.0009019855060614645, + 0.0007950464496389031, + 0.0008937789825722575, + 0.0006242645904421806, + 0.0013062339276075363, + 0.001829802175052464, + 0.002763929544016719, + 0.0013574910117313266, + 0.0014801331562921405, + 0.00520726852118969, + 0.0011808357667177916, + 0.0048593091778457165, + 0.0025482287164777517, + 0.001767785637639463, + 0.005152356810867786, + 0.0022305084858089685, + 0.0028622213285416365, + 0.01324930414557457, + 0.0018466335022822022, + 0.012858995236456394, + 0.033743780106306076, + 0.03294673189520836, + 0.029352372512221336, + 0.020378269255161285, + 0.016749221831560135, + 0.3627270758152008, + 0.0252369474619627, + 0.35451480746269226, + 0.001302618533372879, + 0.002134751295670867, + 0.00325571047142148, + 0.0015975328860804439, + 0.0026787384413182735, + 0.0018505878979340196, + 0.0011654881527647376, + 0.0024054122623056173, + 0.001087840530090034, + 0.001963225658982992, + 0.0011627309722825885, + 0.001357005094178021, + 0.001680584973655641, + 0.0009273673640564084, + 0.0010033162543550134, + 0.000806035881396383, + 0.001094487146474421, + 0.001341949449852109, + 0.0011329733533784747, + 0.000785859243478626, + 0.002455272478982806, + 0.001130233402363956, + 0.0005960604758001864, + 0.0005992415826767683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.08904317021369934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_tensor([8519, 8519])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_tensor([8519, 8519])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..907f401d7e63acbe3237675eb82e67de54b3f16a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_tensor([8519, 8519])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, uninterrupted 360-degree tracking shot captures a pair of emperor angelfish engaged in an elegant courtship ritual, their entire world rendered in the signature style of Hokusai's ukiyo-e woodblock prints. The two fish's bodies, defined by sharp black outlines, are a vibrant composition of flat, unblended colors; their forms are streaked with bold horizontal stripes of vibrant lemon yellow and deep blue. They move in perfect, harmonious sync, swirling fluidly from nose to tail in an elegant underwater ballet. Their fins undulate with a subtle, stylized motion, propelling them through an ocean composed not of water but of graphic forms. The background is a dynamic tapestry of stylized, curling waves, their crests rising into claw-like masses of pure white foam, contrasting sharply against the flat Prussian blue plane of the deep ocean, creating a powerful sense of movement and depth. There is no soft, diffused light; instead, the scene embodies the high-contrast, two-dimensional aesthetic of woodblock prints, where form and color dominate, evoking a profound atmosphere of tranquility, timeless beauty, and perfect natural harmony.", + "seed": 18858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_8519_seed_18858_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23762633707374334, + "vlm_logprob": -0.23762633707374334, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.23762633707374334, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03984266519546509 + ], + "opposite_logprob_tokens": [ + -3.2898426055908203 + ], + "logprob_reward": -0.03984266519546509, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03984266519546509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16627809405326843, + 0.1565188616514206, + 0.08760638535022736, + 0.11221164464950562, + 0.16550156474113464, + 0.0409373939037323, + 0.056385427713394165, + 0.15908865630626678, + 0.02164233848452568, + 0.0338297076523304 + ], + "credit_map": [ + 0.01713475026190281, + 0.015260339714586735, + 0.019881442189216614, + 0.015826541930437088, + 0.03246414661407471, + 0.026234325021505356, + 0.018708376213908195, + 0.020768146961927414, + 0.018418969586491585, + 0.020177850499749184, + 0.013087257742881775, + 0.013432564213871956, + 0.030091581866145134, + 0.026690306141972542, + 0.017072446644306183, + 0.017547862604260445, + 0.007597116753458977, + 0.011068688705563545, + 0.01138207409530878, + 0.010332618840038776, + 0.014418840408325195, + 0.011289019137620926, + 0.010489114560186863, + 0.011028905399143696, + 0.023065119981765747, + 0.01814117096364498, + 0.010881119407713413, + 0.00825238786637783, + 0.01816445402801037, + 0.015400822274386883, + 0.008040224201977253, + 0.010266334749758244, + 0.015367845073342323, + 0.02103673294186592, + 0.018225248903036118, + 0.033443570137023926, + 0.018137307837605476, + 0.014113888144493103, + 0.0191360954195261, + 0.02604086697101593, + 0.006757825147360563, + 0.0057657440192997456, + 0.004592969082295895, + 0.005638204049319029, + 0.004044980276376009, + 0.0032877842895686626, + 0.004973019473254681, + 0.005876864306628704, + 0.01608721911907196, + 0.003614661982282996, + 0.003910123370587826, + 0.0065140253864228725, + 0.00235582422465086, + 0.01585960201919079, + 0.004608883522450924, + 0.0034350885543972254, + 0.014333444647490978, + 0.016026649624109268, + 0.030240138992667198, + 0.016431793570518494, + 0.012868338264524937, + 0.02984992042183876, + 0.018361130729317665, + 0.020977221429347992, + 0.0029494455084204674, + 0.002211179817095399, + 0.002642429433763027, + 0.0031784127932041883, + 0.0029161523561924696, + 0.0028122395742684603, + 0.0025282802525907755, + 0.0024041966535151005, + 0.0038952615577727556, + 0.0032545297872275114, + 0.0055000861175358295, + 0.005488145165145397, + 0.003402609610930085, + 0.0034664678387343884, + 0.005424498580396175, + 0.0033981045708060265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03984266519546509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017269235104322433 + ], + "opposite_logprob_tokens": [ + -6.376727104187012 + ], + "logprob_reward": -0.0017269235104322433, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017269235104322433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14493459463119507, + 0.0771985873579979, + 0.08204639703035355, + 0.11108096688985825, + 0.08393795788288116, + 0.06271474808454514, + 0.1659097969532013, + 0.16138634085655212, + 0.0730695128440857, + 0.03772110491991043 + ], + "credit_map": [ + 0.02610670030117035, + 0.013284604996442795, + 0.02085648663341999, + 0.014319960959255695, + 0.01980796456336975, + 0.013762198388576508, + 0.018634214997291565, + 0.018162470310926437, + 0.01037486083805561, + 0.011135779321193695, + 0.006229955703020096, + 0.007577658165246248, + 0.012793763540685177, + 0.011820324696600437, + 0.01027077715843916, + 0.006995467469096184, + 0.0077639683149755, + 0.014690183103084564, + 0.010034274309873581, + 0.014484531246125698, + 0.01148976106196642, + 0.008019890636205673, + 0.009835031814873219, + 0.005728754680603743, + 0.0225719902664423, + 0.01578489877283573, + 0.009523208253085613, + 0.011824281886219978, + 0.01172710582613945, + 0.010936195962131023, + 0.016473302617669106, + 0.012239976786077023, + 0.008955627679824829, + 0.012811578810214996, + 0.009217793121933937, + 0.015029250644147396, + 0.007859945297241211, + 0.006733877118676901, + 0.010027775540947914, + 0.013302110135555267, + 0.013882150873541832, + 0.00814442802220583, + 0.0057346634566783905, + 0.006134089548140764, + 0.0065259430557489395, + 0.00796419382095337, + 0.006774317007511854, + 0.0075549655593931675, + 0.048562876880168915, + 0.010427997447550297, + 0.015842068940401077, + 0.012315789237618446, + 0.009491411969065666, + 0.050497084856033325, + 0.008404938504099846, + 0.010367625392973423, + 0.012141712941229343, + 0.02259654738008976, + 0.02446592040359974, + 0.010171541944146156, + 0.016406310722231865, + 0.050180237740278244, + 0.011903194710612297, + 0.013520870357751846, + 0.008329653181135654, + 0.008253466337919235, + 0.009387582540512085, + 0.008310290053486824, + 0.007568082306534052, + 0.00897303782403469, + 0.007257263176143169, + 0.014990139752626419, + 0.005270218942314386, + 0.0033415616489946842, + 0.005774694494903088, + 0.004639309365302324, + 0.005166074261069298, + 0.004754128400236368, + 0.0048238784074783325, + 0.00395123939961195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0017269235104322433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26556095480918884, + 0.06120917573571205, + 0.0523967519402504, + 0.08068524301052094, + 0.0612630769610405, + 0.08316970616579056, + 0.08406172692775726, + 0.1869317591190338, + 0.03858790174126625, + 0.08613371849060059 + ], + "credit_map": [ + 0.039945583790540695, + 0.027022095397114754, + 0.022198202088475227, + 0.01836484484374523, + 0.05735005810856819, + 0.040159840136766434, + 0.026694031432271004, + 0.0338263064622879, + 0.007278930861502886, + 0.007305690553039312, + 0.007371851243078709, + 0.007137875072658062, + 0.008500419557094574, + 0.00809647236019373, + 0.007636495400220156, + 0.007881438359618187, + 0.006270097568631172, + 0.0068822139874100685, + 0.005543477367609739, + 0.006486592348664999, + 0.009506776928901672, + 0.005379474721848965, + 0.00685234647244215, + 0.005475773010402918, + 0.01041480153799057, + 0.010041113942861557, + 0.008620794862508774, + 0.008743240498006344, + 0.012225761078298092, + 0.011543777771294117, + 0.011084658093750477, + 0.008011101745069027, + 0.008018947206437588, + 0.008342726156115532, + 0.00650373799726367, + 0.008008171804249287, + 0.008127137087285519, + 0.006269750650972128, + 0.006519176997244358, + 0.009473427198827267, + 0.017404792830348015, + 0.010500767268240452, + 0.007198183331638575, + 0.009344702586531639, + 0.010237469337880611, + 0.009398826397955418, + 0.009394134394824505, + 0.009690831415355206, + 0.024213870987296104, + 0.0057428446598351, + 0.007626088336110115, + 0.006480508949607611, + 0.005692075472325087, + 0.024380167946219444, + 0.004787474405020475, + 0.00513869384303689, + 0.015183490701019764, + 0.012272663414478302, + 0.058852195739746094, + 0.010928220115602016, + 0.011368426494300365, + 0.055378466844558716, + 0.010513685643672943, + 0.012434614822268486, + 0.006032170262187719, + 0.004486073739826679, + 0.0044588069431483746, + 0.004694837611168623, + 0.004547256510704756, + 0.007032574620097876, + 0.003525181906297803, + 0.0038109992165118456, + 0.012065666727721691, + 0.007837939076125622, + 0.008194752037525177, + 0.00861362461000681, + 0.012956595979630947, + 0.015583666041493416, + 0.012299287132918835, + 0.008582185953855515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028160461224615574 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028160461224615574, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028160461224615574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11276377737522125, + 0.08085647225379944, + 0.021980416029691696, + 0.055658746510744095, + 0.036607470363378525, + 0.12069429457187653, + 0.04084664583206177, + 0.4270334541797638, + 0.027545172721147537, + 0.07601362466812134 + ], + "credit_map": [ + 0.01245779637247324, + 0.0116506889462471, + 0.013516145758330822, + 0.011718363501131535, + 0.014903465285897255, + 0.01730094663798809, + 0.016234898939728737, + 0.014981462620198727, + 0.011819109320640564, + 0.010251753963530064, + 0.005989024415612221, + 0.005505937151610851, + 0.0132581926882267, + 0.015576181001961231, + 0.007862047292292118, + 0.010594220831990242, + 0.0024039612617343664, + 0.0027168854139745235, + 0.002782343653962016, + 0.0035327733494341373, + 0.003553888527676463, + 0.002357388148084283, + 0.0027052529621869326, + 0.001927921548485756, + 0.013292456045746803, + 0.0092694740742445, + 0.004002783913165331, + 0.005146498791873455, + 0.007877841591835022, + 0.005843048915266991, + 0.004163977690041065, + 0.006062653381377459, + 0.005208659917116165, + 0.00538115156814456, + 0.005815902259200811, + 0.005006909836083651, + 0.0047669792547822, + 0.003067153040319681, + 0.003044012701138854, + 0.004316696897149086, + 0.023050611838698387, + 0.0162026509642601, + 0.00792534276843071, + 0.013734226115047932, + 0.013183299452066422, + 0.02580876462161541, + 0.010510790161788464, + 0.010278597474098206, + 0.01122367661446333, + 0.0030608531087636948, + 0.0027210500556975603, + 0.00400236900895834, + 0.0020523667335510254, + 0.012323287315666676, + 0.002641615690663457, + 0.002821424277499318, + 0.05424964055418968, + 0.07420829683542252, + 0.057192761451005936, + 0.03499138355255127, + 0.03303641453385353, + 0.12803390622138977, + 0.024331260472536087, + 0.020989764481782913, + 0.002612218027934432, + 0.002130321692675352, + 0.002243913011625409, + 0.002641070634126663, + 0.0035259011201560497, + 0.005825885105878115, + 0.0021220112685114145, + 0.006443850230425596, + 0.013435762375593185, + 0.008404295891523361, + 0.007884940132498741, + 0.008269976824522018, + 0.012419672682881355, + 0.011113463900983334, + 0.009002795442938805, + 0.005482712760567665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0028160461224615574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370102167129517 + ], + "opposite_logprob_tokens": [ + -0.3870101869106293 + ], + "logprob_reward": -1.1370102167129517, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370102167129517, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.26811733841896057, + 0.04294770583510399, + 0.05582902953028679, + 0.11767059564590454, + 0.04637278616428375, + 0.08253699541091919, + 0.09833849966526031, + 0.08605057746171951, + 0.040046170353889465, + 0.16209033131599426 + ], + "credit_map": [ + 0.036147717386484146, + 0.032527558505535126, + 0.013547087088227272, + 0.020865559577941895, + 0.0360855758190155, + 0.06366093456745148, + 0.01712065562605858, + 0.04816223308444023, + 0.006827326025813818, + 0.007727559190243483, + 0.0038122134283185005, + 0.003959252964705229, + 0.005227752961218357, + 0.004568996839225292, + 0.005332087632268667, + 0.0054925186559557915, + 0.005775365047156811, + 0.009148416109383106, + 0.004933370277285576, + 0.006338422652333975, + 0.008240441791713238, + 0.008184791542589664, + 0.005110969301313162, + 0.008097250014543533, + 0.015807626768946648, + 0.015432399697601795, + 0.0123159633949399, + 0.014141673222184181, + 0.017676077783107758, + 0.014207166619598866, + 0.015500671230256557, + 0.012589013203978539, + 0.007315333001315594, + 0.0066599734127521515, + 0.005495530553162098, + 0.006045990623533726, + 0.007837092503905296, + 0.003569171065464616, + 0.003989723045378923, + 0.005459974054247141, + 0.013865283690392971, + 0.012982192449271679, + 0.010174417868256569, + 0.009603526443243027, + 0.009597470052540302, + 0.008447907865047455, + 0.0077467490918934345, + 0.010119440965354443, + 0.02629260905086994, + 0.008127519860863686, + 0.011695469729602337, + 0.007296286057680845, + 0.0061027961783111095, + 0.026411255821585655, + 0.005701051559299231, + 0.006711519323289394, + 0.00864377897232771, + 0.007723371498286724, + 0.01878141611814499, + 0.01007930003106594, + 0.008492758497595787, + 0.01567048579454422, + 0.007760702166706324, + 0.00889876764267683, + 0.006525569595396519, + 0.003524154657498002, + 0.0043267374858260155, + 0.004404845647513866, + 0.005551982205361128, + 0.007284845225512981, + 0.003580282209441066, + 0.004847753793001175, + 0.016999179497361183, + 0.01368188951164484, + 0.016131239011883736, + 0.019292807206511497, + 0.022633058950304985, + 0.033222757279872894, + 0.025089509785175323, + 0.015039884485304356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -1.1370102167129517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, uninterrupted 360-degree tracking shot captures a pair of emperor angelfish engaged in an elegant courtship ritual, their entire world rendered in the signature style of Hokusai's ukiyo-e woodblock prints. The two fish's bodies, defined by sharp black outlines, are a vibrant composition of flat, unblended colors; their forms are streaked with bold horizontal stripes of vibrant lemon yellow and deep blue. They move in perfect, harmonious sync, swirling fluidly from nose to tail in an elegant underwater ballet. Their fins undulate with a subtle, stylized motion, propelling them through an ocean composed not of water but of graphic forms. The background is a dynamic tapestry of stylized, curling waves, their crests rising into claw-like masses of pure white foam, contrasting sharply against the flat Prussian blue plane of the deep ocean, creating a powerful sense of movement and depth. There is no soft, diffused light; instead, the scene embodies the high-contrast, two-dimensional aesthetic of woodblock prints, where form and color dominate, evoking a profound atmosphere of tranquility, timeless beauty, and perfect natural harmony.", + "seed": 18859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_8519_seed_18859_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053008076874539255, + "vlm_logprob": -0.053008076874539255, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.053008076874539255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027877436950802803 + ], + "opposite_logprob_tokens": [ + -3.6528773307800293 + ], + "logprob_reward": -0.027877436950802803, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027877436950802803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2689785063266754, + 0.09865344315767288, + 0.04338911175727844, + 0.06727194041013718, + 0.08094777911901474, + 0.18520641326904297, + 0.06954993307590485, + 0.0356406643986702, + 0.014714239165186882, + 0.1356479376554489 + ], + "credit_map": [ + 0.02217830903828144, + 0.03342308849096298, + 0.022593943402171135, + 0.0316937118768692, + 0.061994168907403946, + 0.0255726408213377, + 0.029458848759531975, + 0.04206377640366554, + 0.011315050534904003, + 0.013370944187045097, + 0.012201255187392235, + 0.011962051503360271, + 0.01191287487745285, + 0.011410634033381939, + 0.013029679656028748, + 0.013450955972075462, + 0.0062789018265903, + 0.00357228796929121, + 0.0040309838950634, + 0.007249763701111078, + 0.003313623368740082, + 0.005989431869238615, + 0.008058605715632439, + 0.004895511083304882, + 0.004225291311740875, + 0.004611376207321882, + 0.00895778276026249, + 0.005749083124101162, + 0.004423457197844982, + 0.007301841396838427, + 0.00932486541569233, + 0.02267823927104473, + 0.011393388733267784, + 0.008536065928637981, + 0.008617192506790161, + 0.013487577438354492, + 0.008690599352121353, + 0.009069322608411312, + 0.013263137079775333, + 0.007890496402978897, + 0.031254589557647705, + 0.02005842514336109, + 0.02009458653628826, + 0.015474972315132618, + 0.02301350235939026, + 0.021342342719435692, + 0.01951959729194641, + 0.03444839268922806, + 0.004398994613438845, + 0.004455328453332186, + 0.0035272755194455385, + 0.0048100207932293415, + 0.021555529907345772, + 0.009774664416909218, + 0.0055520543828606606, + 0.015476072207093239, + 0.005543587263673544, + 0.0029630884528160095, + 0.003919265232980251, + 0.00661092484369874, + 0.003066621022298932, + 0.004891387652605772, + 0.003848972264677286, + 0.00479681883007288, + 0.0020676024723798037, + 0.0012770729372277856, + 0.001831025118008256, + 0.0021902932785451412, + 0.0011868479195982218, + 0.002441627671942115, + 0.0018107889918610454, + 0.001908981124870479, + 0.030088480561971664, + 0.05011250078678131, + 0.005508117843419313, + 0.01169605739414692, + 0.010136330500245094, + 0.010537252761423588, + 0.0073016430251300335, + 0.010267556644976139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.027877436950802803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003204688895493746 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.003204688895493746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204688895493746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20570434629917145, + 0.09307076036930084, + 0.13467812538146973, + 0.182495579123497, + 0.04398536682128906, + 0.06439902633428574, + 0.14047852158546448, + 0.053937457501888275, + 0.02714161016047001, + 0.054109230637550354 + ], + "credit_map": [ + 0.02110774628818035, + 0.04223601147532463, + 0.02282893843948841, + 0.028811300173401833, + 0.027247203513979912, + 0.018854975700378418, + 0.022588593885302544, + 0.0220295712351799, + 0.011473232880234718, + 0.010084335692226887, + 0.01250770129263401, + 0.011385860852897167, + 0.01157025620341301, + 0.008750485256314278, + 0.01177613902837038, + 0.015522748231887817, + 0.016450168564915657, + 0.012786313891410828, + 0.01521337777376175, + 0.012188197113573551, + 0.011735768988728523, + 0.03131454437971115, + 0.014501114375889301, + 0.020488649606704712, + 0.016901716589927673, + 0.007952116429805756, + 0.033201467245817184, + 0.009241892956197262, + 0.008276755921542645, + 0.02641393058001995, + 0.012993480078876019, + 0.06751421093940735, + 0.007160243112593889, + 0.004245862830430269, + 0.004244969226419926, + 0.006735931616276503, + 0.005464150104671717, + 0.00601175706833601, + 0.004458627663552761, + 0.005663823336362839, + 0.006750623229891062, + 0.007988912984728813, + 0.007489461917430162, + 0.005985603667795658, + 0.009595941752195358, + 0.009481680579483509, + 0.0096916314214468, + 0.007415170315653086, + 0.010083830915391445, + 0.010669118724763393, + 0.00892980583012104, + 0.006781414151191711, + 0.036907825618982315, + 0.022063281387090683, + 0.009023167192935944, + 0.036020077764987946, + 0.008472534827888012, + 0.00597346480935812, + 0.006010942626744509, + 0.007702965755015612, + 0.006140158511698246, + 0.006006017792969942, + 0.005446773488074541, + 0.008184600621461868, + 0.0032235251273959875, + 0.00204869220033288, + 0.004894434008747339, + 0.0028341077268123627, + 0.002533724531531334, + 0.003537427634000778, + 0.0036616830620914698, + 0.00440801540389657, + 0.007389141712337732, + 0.012064578011631966, + 0.0036105485633015633, + 0.008033045567572117, + 0.005320933647453785, + 0.006928390357643366, + 0.005127425771206617, + 0.005635168869048357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.003204688895493746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947985220700502 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947985220700502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947985220700502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14964470267295837, + 0.03781164437532425, + 0.03092588670551777, + 0.028166739270091057, + 0.014729522168636322, + 0.516534686088562, + 0.08643893897533417, + 0.02487962320446968, + 0.022406520321965218, + 0.08846170455217361 + ], + "credit_map": [ + 0.01211092434823513, + 0.01916825771331787, + 0.01260348130017519, + 0.020018137991428375, + 0.03672122582793236, + 0.012673372402787209, + 0.015576925128698349, + 0.020772378891706467, + 0.0040473779663443565, + 0.0053212689235806465, + 0.004398400895297527, + 0.005461311899125576, + 0.004998435731977224, + 0.004272771999239922, + 0.004285137169063091, + 0.0050269439816474915, + 0.0038815161678940058, + 0.0028167462442070246, + 0.0034063588827848434, + 0.0037657495122402906, + 0.0022908439859747887, + 0.005441790912300348, + 0.004580737091600895, + 0.004742146469652653, + 0.0047149620950222015, + 0.002068080473691225, + 0.005725205410271883, + 0.0018919683061540127, + 0.0021931291557848454, + 0.003285544691607356, + 0.002648883732035756, + 0.005638969596475363, + 0.002004271373152733, + 0.0015827311435714364, + 0.0012905318289995193, + 0.002497190609574318, + 0.001842275494709611, + 0.0019780548755079508, + 0.0016280477866530418, + 0.0019064197549596429, + 0.08514050394296646, + 0.05530637502670288, + 0.05473781377077103, + 0.04427064210176468, + 0.06574668735265732, + 0.052790671586990356, + 0.056128766387701035, + 0.10241326689720154, + 0.005932303611189127, + 0.0061482698656618595, + 0.004604243207722902, + 0.004760615061968565, + 0.026219407096505165, + 0.013259392231702805, + 0.0055960374884307384, + 0.01991867832839489, + 0.005138194654136896, + 0.002369871363043785, + 0.003459833562374115, + 0.0035308923106640577, + 0.0020950629841536283, + 0.0033622276969254017, + 0.002299480140209198, + 0.0026240639854222536, + 0.0021239782217890024, + 0.0021350369788706303, + 0.0019505374366417527, + 0.003349719801917672, + 0.0020416590850800276, + 0.002782070077955723, + 0.004513060208410025, + 0.003510460490360856, + 0.014524046331644058, + 0.03377654775977135, + 0.0026785621885210276, + 0.006373608950525522, + 0.0067296382039785385, + 0.0076324669644236565, + 0.00433736527338624, + 0.012409471906721592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.005947985220700502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024862352292984724 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.0024862352292984724, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024862352292984724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14703527092933655, + 0.07135404646396637, + 0.03875723481178284, + 0.041001539677381516, + 0.021152792498469353, + 0.37944865226745605, + 0.11666838824748993, + 0.022573376074433327, + 0.019715091213583946, + 0.14229357242584229 + ], + "credit_map": [ + 0.012519032694399357, + 0.018182620406150818, + 0.014544700272381306, + 0.020964249968528748, + 0.03367593139410019, + 0.011845295317471027, + 0.01533515751361847, + 0.019968295469880104, + 0.007227452006191015, + 0.008057250641286373, + 0.008853154256939888, + 0.01108639407902956, + 0.00916778203099966, + 0.007836300879716873, + 0.009747911244630814, + 0.00937781110405922, + 0.004430803004652262, + 0.0036861561238765717, + 0.00524721248075366, + 0.005722930189222097, + 0.003174319863319397, + 0.006004318129271269, + 0.006200899835675955, + 0.004290597978979349, + 0.0021563235204666853, + 0.0024520307779312134, + 0.0052654799073934555, + 0.002167694503441453, + 0.0020261043682694435, + 0.008322669193148613, + 0.0020699312444776297, + 0.016541309654712677, + 0.003221452236175537, + 0.002120508113875985, + 0.0021844059228897095, + 0.0035378148313611746, + 0.0028515602461993694, + 0.0026308451779186726, + 0.002326271263882518, + 0.0022799342405050993, + 0.062496647238731384, + 0.04061184450984001, + 0.04059852659702301, + 0.03243272006511688, + 0.048761557787656784, + 0.03881701081991196, + 0.040989115834236145, + 0.07474122941493988, + 0.008084188215434551, + 0.007234647870063782, + 0.0065553761087358, + 0.004656923469156027, + 0.036826152354478836, + 0.018005039542913437, + 0.007284127175807953, + 0.02802194282412529, + 0.0034222183749079704, + 0.002318775048479438, + 0.002154748188331723, + 0.0037915720604360104, + 0.002392499940469861, + 0.002719487063586712, + 0.002522624097764492, + 0.0032514543272554874, + 0.0020292706321924925, + 0.00144676654599607, + 0.0027869304176419973, + 0.002657623030245304, + 0.0017508192686364055, + 0.002325566252693534, + 0.0033692142460495234, + 0.0033489037305116653, + 0.025398053228855133, + 0.053254105150699615, + 0.004774985369294882, + 0.01224821899086237, + 0.011039740405976772, + 0.012175302021205425, + 0.007169247604906559, + 0.016233932226896286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0024862352292984724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22552403807640076 + ], + "opposite_logprob_tokens": [ + -1.6005240678787231 + ], + "logprob_reward": -0.22552403807640076, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22552403807640076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06670694053173065, + 0.07489436119794846, + 0.027810046449303627, + 0.043985866010189056, + 0.032401103526353836, + 0.48704794049263, + 0.11498881876468658, + 0.015049338340759277, + 0.022671084851026535, + 0.11444448679685593 + ], + "credit_map": [ + 0.007272784598171711, + 0.011108817532658577, + 0.005455544218420982, + 0.007688031531870365, + 0.011149769648909569, + 0.008853152394294739, + 0.007886088453233242, + 0.007292757276445627, + 0.007934167049825191, + 0.005020055919885635, + 0.010462048463523388, + 0.01108874287456274, + 0.010380709543824196, + 0.00705678341910243, + 0.009462320245802402, + 0.01348953228443861, + 0.004098997917026281, + 0.0021540061570703983, + 0.004224186763167381, + 0.002967692678794265, + 0.0018674961756914854, + 0.00514871533960104, + 0.0030602416954934597, + 0.004288710653781891, + 0.004910778719931841, + 0.0028110335115343332, + 0.0071912165731191635, + 0.002469538478180766, + 0.0028552343137562275, + 0.007749741896986961, + 0.0032175304368138313, + 0.012780792079865932, + 0.005014047026634216, + 0.003121407004073262, + 0.002383264945819974, + 0.006139139179140329, + 0.004950134549289942, + 0.003277386073023081, + 0.003432936267927289, + 0.004082788713276386, + 0.08043847978115082, + 0.05214734375476837, + 0.05157933011651039, + 0.042098138481378555, + 0.061292167752981186, + 0.050185929983854294, + 0.0534847155213356, + 0.0958217978477478, + 0.007513702381402254, + 0.008491702377796173, + 0.00547364167869091, + 0.005303678568452597, + 0.03809511289000511, + 0.017524272203445435, + 0.006854660343378782, + 0.025732047855854034, + 0.002063784981146455, + 0.001372242346405983, + 0.0018117072759196162, + 0.0032689557410776615, + 0.0011455556377768517, + 0.0016749162459746003, + 0.0019229884492233396, + 0.001789188594557345, + 0.0031267968006432056, + 0.0025455451104789972, + 0.0032437953632324934, + 0.00237518223002553, + 0.002014446770772338, + 0.0031700031831860542, + 0.003000801894813776, + 0.003194513963535428, + 0.024808302521705627, + 0.04304763674736023, + 0.004348463844507933, + 0.010857876390218735, + 0.008474331349134445, + 0.009150110185146332, + 0.00533067574724555, + 0.008427086286246777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.22552403807640076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, uninterrupted 360-degree tracking shot captures a pair of emperor angelfish engaged in an elegant courtship ritual, their entire world rendered in the signature style of Hokusai's ukiyo-e woodblock prints. The two fish's bodies, defined by sharp black outlines, are a vibrant composition of flat, unblended colors; their forms are streaked with bold horizontal stripes of vibrant lemon yellow and deep blue. They move in perfect, harmonious sync, swirling fluidly from nose to tail in an elegant underwater ballet. Their fins undulate with a subtle, stylized motion, propelling them through an ocean composed not of water but of graphic forms. The background is a dynamic tapestry of stylized, curling waves, their crests rising into claw-like masses of pure white foam, contrasting sharply against the flat Prussian blue plane of the deep ocean, creating a powerful sense of movement and depth. There is no soft, diffused light; instead, the scene embodies the high-contrast, two-dimensional aesthetic of woodblock prints, where form and color dominate, evoking a profound atmosphere of tranquility, timeless beauty, and perfect natural harmony.", + "seed": 18860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_8519_seed_18860_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2064330353634432, + "vlm_logprob": -0.2064330353634432, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.2064330353634432, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040080562233924866 + ], + "opposite_logprob_tokens": [ + -3.2900805473327637 + ], + "logprob_reward": -0.040080562233924866, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.040080562233924866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15016889572143555, + 0.09903813898563385, + 0.2440362423658371, + 0.10986053198575974, + 0.07314196974039078, + 0.0914146900177002, + 0.0617702417075634, + 0.03781619668006897, + 0.061217039823532104, + 0.07153608649969101 + ], + "credit_map": [ + 0.019624939188361168, + 0.029572436586022377, + 0.012802882120013237, + 0.022587750107049942, + 0.01446433924138546, + 0.024435071274638176, + 0.01469961367547512, + 0.011981848627328873, + 0.012180212885141373, + 0.013252691365778446, + 0.008509312756359577, + 0.013562032021582127, + 0.016364814713597298, + 0.014373891055583954, + 0.01011169608682394, + 0.010683488100767136, + 0.03686824440956116, + 0.028590917587280273, + 0.01628648303449154, + 0.023873116821050644, + 0.05626845359802246, + 0.036027733236551285, + 0.025538086891174316, + 0.020583176985383034, + 0.014696113765239716, + 0.01088723260909319, + 0.011683150194585323, + 0.013401651754975319, + 0.020107652992010117, + 0.0157772246748209, + 0.012777687050402164, + 0.010529817081987858, + 0.010909879580140114, + 0.009206721559166908, + 0.006372862029820681, + 0.007751255761831999, + 0.012731365859508514, + 0.008920499123632908, + 0.01078470703214407, + 0.006464672740548849, + 0.009278163313865662, + 0.02818562649190426, + 0.01238330826163292, + 0.0062417443841695786, + 0.012091165408492088, + 0.010299567133188248, + 0.004854485392570496, + 0.00808061845600605, + 0.006322438828647137, + 0.007419315632432699, + 0.005131907295435667, + 0.007075670175254345, + 0.008440644480288029, + 0.008374184370040894, + 0.0068703265860676765, + 0.012135746888816357, + 0.0027843716088682413, + 0.004119371995329857, + 0.0037058850284665823, + 0.0038491638842970133, + 0.007410758640617132, + 0.0047974977642297745, + 0.006943823769688606, + 0.00420532189309597, + 0.005323654506355524, + 0.004870926029980183, + 0.006137555930763483, + 0.0054068309254944324, + 0.0139915207400918, + 0.008414920419454575, + 0.01117568090558052, + 0.0058959475718438625, + 0.007863443344831467, + 0.007367598824203014, + 0.007779582869261503, + 0.008337855339050293, + 0.015664372593164444, + 0.008209774270653725, + 0.008648310787975788, + 0.007665146607905626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.040080562233924866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017253764672204852 + ], + "opposite_logprob_tokens": [ + -6.376725196838379 + ], + "logprob_reward": -0.0017253764672204852, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017253764672204852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1453464776277542, + 0.11333850026130676, + 0.12241819500923157, + 0.05073937028646469, + 0.05226597934961319, + 0.15938769280910492, + 0.06073505058884621, + 0.18315809965133667, + 0.07407883554697037, + 0.03853178396821022 + ], + "credit_map": [ + 0.02522583305835724, + 0.020844878628849983, + 0.014959077350795269, + 0.009209688752889633, + 0.03424481302499771, + 0.019934721291065216, + 0.010931629687547684, + 0.009995832107961178, + 0.014626418240368366, + 0.0102064935490489, + 0.010446949861943722, + 0.01483816560357809, + 0.014007041230797768, + 0.014967869035899639, + 0.011527049355208874, + 0.022718513384461403, + 0.02003621496260166, + 0.015368043445050716, + 0.01212336216121912, + 0.009882412850856781, + 0.02970537357032299, + 0.016247274354100227, + 0.010477649047970772, + 0.008577865548431873, + 0.006524617318063974, + 0.005805433262139559, + 0.007970315404236317, + 0.007734350394457579, + 0.00583948940038681, + 0.006698041222989559, + 0.005132105201482773, + 0.005035016685724258, + 0.010004754178225994, + 0.006488265935331583, + 0.006276935804635286, + 0.004895585589110851, + 0.00593603728339076, + 0.007477167062461376, + 0.006272278260439634, + 0.004914955236017704, + 0.02214141935110092, + 0.010783982463181019, + 0.0423535518348217, + 0.011095256544649601, + 0.009474020451307297, + 0.01763446070253849, + 0.011245936155319214, + 0.03465906158089638, + 0.007256034296005964, + 0.006470825523138046, + 0.0062195551581680775, + 0.00761522026732564, + 0.007764095440506935, + 0.008407799527049065, + 0.00652791652828455, + 0.010473610833287239, + 0.022922324016690254, + 0.030788060277700424, + 0.01774211786687374, + 0.014486879110336304, + 0.02757113054394722, + 0.02300240471959114, + 0.02659536525607109, + 0.02004982903599739, + 0.008134428411722183, + 0.008354242891073227, + 0.010209317319095135, + 0.008473473601043224, + 0.010944250971078873, + 0.00956717412918806, + 0.010637152940034866, + 0.007758795283734798, + 0.004081699065864086, + 0.004171603359282017, + 0.003214676631614566, + 0.003720944980159402, + 0.00988587737083435, + 0.004720689263194799, + 0.004551687743514776, + 0.004184604622423649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0017253764672204852 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253500021994114 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253500021994114, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253500021994114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2124742716550827, + 0.06490366905927658, + 0.10591354966163635, + 0.08011788874864578, + 0.11162947863340378, + 0.1047782450914383, + 0.04221930354833603, + 0.11741792410612106, + 0.06941094994544983, + 0.09113472700119019 + ], + "credit_map": [ + 0.02836673893034458, + 0.04852650687098503, + 0.013739976100623608, + 0.023563213646411896, + 0.03890237584710121, + 0.035338327288627625, + 0.013198579661548138, + 0.010838565416634083, + 0.007948899641633034, + 0.008028779178857803, + 0.005658945068717003, + 0.007465359754860401, + 0.006869954522699118, + 0.01128240767866373, + 0.007071374915540218, + 0.010577945969998837, + 0.015468957833945751, + 0.011595440097153187, + 0.01588604785501957, + 0.009084383025765419, + 0.014508136548101902, + 0.01542193628847599, + 0.0131690613925457, + 0.010779580101370811, + 0.009766322560608387, + 0.006956157740205526, + 0.010815935209393501, + 0.013002853840589523, + 0.010953392833471298, + 0.010480047203600407, + 0.010528726503252983, + 0.007614457048475742, + 0.019143208861351013, + 0.013920947909355164, + 0.011085051111876965, + 0.011678279377520084, + 0.020947933197021484, + 0.013024888001382351, + 0.012681130319833755, + 0.009148037061095238, + 0.01213506143540144, + 0.014735064469277859, + 0.030347008258104324, + 0.006260600872337818, + 0.00793154165148735, + 0.011215324513614178, + 0.007012550253421068, + 0.015141090378165245, + 0.005711218807846308, + 0.005050574894994497, + 0.004704743158072233, + 0.004431823268532753, + 0.005566843785345554, + 0.005331684835255146, + 0.0063041565008461475, + 0.005118257366120815, + 0.011300358921289444, + 0.012830806896090508, + 0.01435128878802061, + 0.021974142640829086, + 0.012686895206570625, + 0.0141085060313344, + 0.013617593795061111, + 0.01654832623898983, + 0.006176060065627098, + 0.006584700662642717, + 0.008460872806608677, + 0.008951880037784576, + 0.014451657421886921, + 0.008274250663816929, + 0.00967399775981903, + 0.006837530992925167, + 0.006698446348309517, + 0.010614292696118355, + 0.007343670818954706, + 0.007649075239896774, + 0.022078413516283035, + 0.010481368750333786, + 0.01601686142385006, + 0.010252595879137516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611591411754489 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.003611591411754489, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611591411754489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19163045287132263, + 0.0954388752579689, + 0.06027858704328537, + 0.15500660240650177, + 0.044737014919519424, + 0.20095445215702057, + 0.06427902728319168, + 0.09461583942174911, + 0.05029122158885002, + 0.042768023908138275 + ], + "credit_map": [ + 0.023356974124908447, + 0.04802343621850014, + 0.011021438054740429, + 0.025572234764695168, + 0.019359048455953598, + 0.04004975035786629, + 0.0119027653709054, + 0.012344802729785442, + 0.013502015732228756, + 0.00992549117654562, + 0.009803859516978264, + 0.012097544968128204, + 0.00798862986266613, + 0.015662213787436485, + 0.014946269802749157, + 0.011512843891978264, + 0.008043930865824223, + 0.006351640447974205, + 0.008038296364247799, + 0.004314087796956301, + 0.009294522926211357, + 0.007882141508162022, + 0.005700765643268824, + 0.010653199627995491, + 0.016904747113585472, + 0.01262813713401556, + 0.02390149049460888, + 0.03168421983718872, + 0.01139482855796814, + 0.02249583788216114, + 0.015516198240220547, + 0.02048114314675331, + 0.006886080838739872, + 0.005849950481206179, + 0.004798444919288158, + 0.005297372117638588, + 0.0055562155321240425, + 0.007398632820695639, + 0.005424882750958204, + 0.0035254312679171562, + 0.022462181746959686, + 0.03938379883766174, + 0.01712854392826557, + 0.022851161658763885, + 0.020344039425253868, + 0.02167361229658127, + 0.022405045107007027, + 0.03470604866743088, + 0.007898027077317238, + 0.00674534123390913, + 0.007110060192644596, + 0.006505141034722328, + 0.007684749085456133, + 0.009296618402004242, + 0.009025776758790016, + 0.010013306513428688, + 0.008459631353616714, + 0.011237294413149357, + 0.011966262944042683, + 0.017018837854266167, + 0.012711203657090664, + 0.011998055502772331, + 0.010377276688814163, + 0.01084726769477129, + 0.004177362658083439, + 0.005106701049953699, + 0.005992949940264225, + 0.006057501304894686, + 0.010481496341526508, + 0.005948334001004696, + 0.008374048396945, + 0.004152821842581034, + 0.006150499451905489, + 0.005426827818155289, + 0.004868984222412109, + 0.005253912415355444, + 0.00670625502243638, + 0.005130319856107235, + 0.004027761053293943, + 0.005203459411859512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.003611591411754489 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.974212646484375 + ], + "opposite_logprob_tokens": [ + -0.4742126166820526 + ], + "logprob_reward": -0.974212646484375, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.974212646484375, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12174351513385773, + 0.07939635962247849, + 0.10520462691783905, + 0.04122697189450264, + 0.04068781062960625, + 0.09731939435005188, + 0.0929655060172081, + 0.04986570030450821, + 0.30256035923957825, + 0.0690297931432724 + ], + "credit_map": [ + 0.015161046758294106, + 0.034551139920949936, + 0.006194189190864563, + 0.016056306660175323, + 0.012053069658577442, + 0.020551735535264015, + 0.007850741036236286, + 0.00932528730481863, + 0.00922098197042942, + 0.00733441486954689, + 0.00768491392955184, + 0.010997039265930653, + 0.008747395128011703, + 0.010500540025532246, + 0.012773438356816769, + 0.012137639336287975, + 0.01700863242149353, + 0.014887239784002304, + 0.008245338685810566, + 0.008617658168077469, + 0.024332964792847633, + 0.012897754088044167, + 0.010708855465054512, + 0.008506184443831444, + 0.006355006713420153, + 0.004346740897744894, + 0.0040540145710110664, + 0.005308740306645632, + 0.005219593178480864, + 0.0059981197118759155, + 0.0036961839068681, + 0.006248574238270521, + 0.005049313884228468, + 0.003937490284442902, + 0.0039194668643176556, + 0.0041874852031469345, + 0.004376341123133898, + 0.006793757434934378, + 0.004239888396114111, + 0.008184066973626614, + 0.007640496827661991, + 0.022709982469677925, + 0.011132558807730675, + 0.007703273091465235, + 0.012327783741056919, + 0.01271910686045885, + 0.008848758414387703, + 0.01423743087798357, + 0.009010911919176579, + 0.007579357363283634, + 0.0077340020798146725, + 0.013484310358762741, + 0.010349645279347897, + 0.01111404038965702, + 0.014055647887289524, + 0.019637593999505043, + 0.0061138649471104145, + 0.005854217801243067, + 0.004540721885859966, + 0.0044107455760240555, + 0.008418968878686428, + 0.006715449504554272, + 0.007209962699562311, + 0.006601764354854822, + 0.024161336943507195, + 0.02868519350886345, + 0.0312555655837059, + 0.03969734534621239, + 0.05380247160792351, + 0.03570469468832016, + 0.06132912635803223, + 0.027924584224820137, + 0.01014605164527893, + 0.008386967703700066, + 0.0063401092775166035, + 0.00701741361990571, + 0.013729029335081577, + 0.00787477195262909, + 0.008110581897199154, + 0.007424867711961269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.974212646484375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, uninterrupted 360-degree tracking shot captures a pair of emperor angelfish engaged in an elegant courtship ritual, their entire world rendered in the signature style of Hokusai's ukiyo-e woodblock prints. The two fish's bodies, defined by sharp black outlines, are a vibrant composition of flat, unblended colors; their forms are streaked with bold horizontal stripes of vibrant lemon yellow and deep blue. They move in perfect, harmonious sync, swirling fluidly from nose to tail in an elegant underwater ballet. Their fins undulate with a subtle, stylized motion, propelling them through an ocean composed not of water but of graphic forms. The background is a dynamic tapestry of stylized, curling waves, their crests rising into claw-like masses of pure white foam, contrasting sharply against the flat Prussian blue plane of the deep ocean, creating a powerful sense of movement and depth. There is no soft, diffused light; instead, the scene embodies the high-contrast, two-dimensional aesthetic of woodblock prints, where form and color dominate, evoking a profound atmosphere of tranquility, timeless beauty, and perfect natural harmony.", + "seed": 18861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_5/index_8519_seed_18861_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.30578522230498495, + "vlm_logprob": -0.30578522230498495, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.30578522230498495, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017259402200579643 + ], + "opposite_logprob_tokens": [ + -4.14225959777832 + ], + "logprob_reward": -0.017259402200579643, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017259402200579643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052643220871686935, + 0.04362231120467186, + 0.02269446849822998, + 0.05752692371606827, + 0.026545412838459015, + 0.4977860748767853, + 0.061614297330379486, + 0.16887064278125763, + 0.016596488654613495, + 0.052100155502557755 + ], + "credit_map": [ + 0.009572046808898449, + 0.0048086498863995075, + 0.005029025487601757, + 0.003834205213934183, + 0.009339852258563042, + 0.00632309028878808, + 0.00787403155118227, + 0.005862318444997072, + 0.0038604736328125, + 0.005917300004512072, + 0.006880975794047117, + 0.005484520457684994, + 0.002874787198379636, + 0.0040801395662128925, + 0.010160494595766068, + 0.004363615997135639, + 0.0030154231935739517, + 0.0016391974641010165, + 0.004373510368168354, + 0.0021495046094059944, + 0.0024676737375557423, + 0.002281362423673272, + 0.0029680756852030754, + 0.003799720201641321, + 0.006166380830109119, + 0.002035886747762561, + 0.005512656643986702, + 0.005111408419907093, + 0.0032278583385050297, + 0.005849028937518597, + 0.022814560681581497, + 0.0068091414868831635, + 0.0036052823998034, + 0.00243068253621459, + 0.003065973985940218, + 0.0028238785453140736, + 0.004354499280452728, + 0.0030011243652552366, + 0.003461001208052039, + 0.003802970051765442, + 0.055602237582206726, + 0.051013391464948654, + 0.04485546052455902, + 0.07111217081546783, + 0.09358357638120651, + 0.04965423047542572, + 0.08095455169677734, + 0.05101048946380615, + 0.009524461813271046, + 0.00968074519187212, + 0.006339240353554487, + 0.008986310102045536, + 0.008175873197615147, + 0.006894543766975403, + 0.00650123180821538, + 0.005511891096830368, + 0.010851334780454636, + 0.012787044048309326, + 0.06683462858200073, + 0.007168618496507406, + 0.005440645385533571, + 0.006171300075948238, + 0.005423011723905802, + 0.05419405922293663, + 0.00134392105974257, + 0.002798546338453889, + 0.002375835319980979, + 0.0018362257396802306, + 0.0028336262330412865, + 0.0017121402779594064, + 0.0012692112941294909, + 0.002426980994641781, + 0.007520924787968397, + 0.005377821624279022, + 0.008634344674646854, + 0.006038859486579895, + 0.00587005028501153, + 0.007385358214378357, + 0.004103753715753555, + 0.007169043645262718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.017259402200579643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250038574449718 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.00250038574449718, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250038574449718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1226690337061882, + 0.0925477147102356, + 0.03967905044555664, + 0.06278935819864273, + 0.0912659615278244, + 0.1949857473373413, + 0.13602206110954285, + 0.17097605764865875, + 0.04946226254105568, + 0.03960276022553444 + ], + "credit_map": [ + 0.023469936102628708, + 0.010948854498565197, + 0.008547189645469189, + 0.008028137497603893, + 0.03618736192584038, + 0.016084328293800354, + 0.011377833783626556, + 0.008025402203202248, + 0.009615358896553516, + 0.018850399181246758, + 0.015448980033397675, + 0.011052480898797512, + 0.008859321475028992, + 0.01093667559325695, + 0.01140010729432106, + 0.006384388078004122, + 0.004460322204977274, + 0.0026638652198016644, + 0.00604706397280097, + 0.002829926321282983, + 0.00518027925863862, + 0.005290337838232517, + 0.006276288069784641, + 0.006930967792868614, + 0.0067894537933170795, + 0.002223386662080884, + 0.00681614875793457, + 0.005118012893944979, + 0.0035228256601840258, + 0.007075235713273287, + 0.024700291454792023, + 0.006544003263115883, + 0.013400444760918617, + 0.004968973807990551, + 0.010689564980566502, + 0.006791230291128159, + 0.010893217287957668, + 0.010018005035817623, + 0.02435094304382801, + 0.010153581388294697, + 0.02278328873217106, + 0.019376937299966812, + 0.01829766482114792, + 0.027433641254901886, + 0.036697205156087875, + 0.019137654453516006, + 0.02924935147166252, + 0.02201000414788723, + 0.015658536925911903, + 0.018841521814465523, + 0.01620972529053688, + 0.019525974988937378, + 0.02106204256415367, + 0.015476681292057037, + 0.015011443756520748, + 0.014236130751669407, + 0.012603354640305042, + 0.01307364460080862, + 0.06599915772676468, + 0.007140366826206446, + 0.003832215443253517, + 0.006154066417366266, + 0.005328022409230471, + 0.05684524402022362, + 0.00563396792858839, + 0.008584694005548954, + 0.007220009341835976, + 0.005317624658346176, + 0.006520845927298069, + 0.006039730738848448, + 0.0028254708740860224, + 0.007319917902350426, + 0.006468754727393389, + 0.00419910904020071, + 0.008764584548771381, + 0.004276323597878218, + 0.005053305998444557, + 0.00344944023527205, + 0.002543639624491334, + 0.0048476047813892365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.00250038574449718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637557547539473 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637557547539473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637557547539473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17633305490016937, + 0.14768613874912262, + 0.025536982342600822, + 0.06645090878009796, + 0.24572087824344635, + 0.11042122542858124, + 0.027803676202893257, + 0.16816675662994385, + 0.008833206258714199, + 0.02304719388484955 + ], + "credit_map": [ + 0.033874716609716415, + 0.013139700517058372, + 0.01158042624592781, + 0.008511915802955627, + 0.04786029830574989, + 0.02400236390531063, + 0.024458756670355797, + 0.012904863804578781, + 0.012751603499054909, + 0.01970628648996353, + 0.014182495884597301, + 0.014088096097111702, + 0.011434071697294712, + 0.01525087933987379, + 0.04526342451572418, + 0.015009270049631596, + 0.003212772076949477, + 0.0022810231894254684, + 0.00394474808126688, + 0.00283128023147583, + 0.004055664408951998, + 0.0026853953022509813, + 0.0034066273365169764, + 0.003119469853118062, + 0.007271063048392534, + 0.0025313040241599083, + 0.0060163685120642185, + 0.005952312145382166, + 0.004140931647270918, + 0.006631976459175348, + 0.026484757661819458, + 0.007422195747494698, + 0.03876486048102379, + 0.021123046055436134, + 0.029476845636963844, + 0.025057105347514153, + 0.04323970898985863, + 0.027216024696826935, + 0.03241169825196266, + 0.02843157947063446, + 0.013639467768371105, + 0.011716640554368496, + 0.010347624309360981, + 0.01488170213997364, + 0.01869831047952175, + 0.01060644630342722, + 0.020268501713871956, + 0.010262524709105492, + 0.0031120909843593836, + 0.004146053921431303, + 0.0023289781529456377, + 0.004157119430601597, + 0.004735888913273811, + 0.002755495021119714, + 0.004006199073046446, + 0.002561850007623434, + 0.011883878149092197, + 0.014272153377532959, + 0.06683751940727234, + 0.006994590163230896, + 0.003939160145819187, + 0.0060357325710356236, + 0.004831960424780846, + 0.05337175354361534, + 0.0010750218061730266, + 0.0013347771018743515, + 0.0013699220726266503, + 0.0011028695153072476, + 0.0011870706221088767, + 0.0009465016191825271, + 0.000843370973598212, + 0.0009736719657666981, + 0.0025174389593303204, + 0.0027228984981775284, + 0.0035231539513915777, + 0.002345220884308219, + 0.004184417892247438, + 0.0028317535761743784, + 0.0018492903327569366, + 0.003073016880080104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.004637557547539473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10540373623371124, + 0.060185324400663376, + 0.07983562350273132, + 0.039779141545295715, + 0.08135385066270828, + 0.13993872702121735, + 0.04485444352030754, + 0.40964099764823914, + 0.02106105349957943, + 0.017947113141417503 + ], + "credit_map": [ + 0.011583499610424042, + 0.010746907442808151, + 0.007473442237824202, + 0.009067339822649956, + 0.009911615401506424, + 0.01453544944524765, + 0.020199481397867203, + 0.021886000409722328, + 0.009888885542750359, + 0.007613178342580795, + 0.012613401748239994, + 0.005985511001199484, + 0.0053430842235684395, + 0.007139573339372873, + 0.003697263542562723, + 0.007904425263404846, + 0.011332323774695396, + 0.008437848649919033, + 0.009528077207505703, + 0.0053779007866978645, + 0.01263405755162239, + 0.008998633362352848, + 0.013966980390250683, + 0.009559801779687405, + 0.004334611352533102, + 0.001395080704241991, + 0.005195791833102703, + 0.003364335745573044, + 0.002588521456345916, + 0.005477086640894413, + 0.013786477968096733, + 0.003637234913185239, + 0.013570389710366726, + 0.006813314277678728, + 0.009922093711793423, + 0.00823412649333477, + 0.01309895422309637, + 0.00905024167150259, + 0.010556075721979141, + 0.010108653455972672, + 0.01591556891798973, + 0.01396642904728651, + 0.012383172288537025, + 0.020394928753376007, + 0.026945697143673897, + 0.014667685143649578, + 0.022701183333992958, + 0.012964063324034214, + 0.006434351205825806, + 0.008712167851626873, + 0.004459379240870476, + 0.006423001177608967, + 0.005473427940160036, + 0.005060823634266853, + 0.003657680470496416, + 0.004633611999452114, + 0.030534952878952026, + 0.034115418791770935, + 0.15914520621299744, + 0.0167543925344944, + 0.00865089986473322, + 0.01454070582985878, + 0.011919251643121243, + 0.13398021459579468, + 0.0020485229324549437, + 0.003173548262566328, + 0.003847468877211213, + 0.002199445851147175, + 0.0024990031961351633, + 0.002739598276093602, + 0.0016641344409435987, + 0.0028893311973661184, + 0.0021388567984104156, + 0.0017260252498090267, + 0.0027251490391790867, + 0.0022425150964409113, + 0.0019765410106629133, + 0.0029861635994166136, + 0.001712574390694499, + 0.002439287956804037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0024867109023034573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two emperor angelfish visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fish display sharp black outlines and flat, unblended colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fish bodies streaked with bold horizontal stripes of lemon yellow and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of stylized, curling waves with white foam crests against a Prussian blue plane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a slow, uninterrupted 360-degree tracking shot around the fish?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020420551300049 + ], + "opposite_logprob_tokens": [ + -0.25204208493232727 + ], + "logprob_reward": -1.5020420551300049, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020420551300049, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.055267442017793655, + 0.04801522567868233, + 0.02940027229487896, + 0.07457657158374786, + 0.034999772906303406, + 0.30290138721466064, + 0.08117633312940598, + 0.3118990659713745, + 0.019813725724816322, + 0.04195021092891693 + ], + "credit_map": [ + 0.008173976093530655, + 0.005553212016820908, + 0.004014946054667234, + 0.004893609322607517, + 0.008692138828337193, + 0.007791860029101372, + 0.004886106122285128, + 0.011261585168540478, + 0.0053425999358296394, + 0.005852189846336842, + 0.007845633663237095, + 0.00515720434486866, + 0.0034798334818333387, + 0.00434682285413146, + 0.010691375471651554, + 0.0052995639853179455, + 0.004187137354165316, + 0.0028779562562704086, + 0.00345997023396194, + 0.0023158667609095573, + 0.005200965795665979, + 0.0033016237430274487, + 0.0021386498119682074, + 0.005918099079281092, + 0.007831976749002934, + 0.0025097483303397894, + 0.010676872916519642, + 0.00617303466424346, + 0.004781745374202728, + 0.009090651758015156, + 0.026881886646151543, + 0.006630652118474245, + 0.0033293101005256176, + 0.0026426436379551888, + 0.0035040932707488537, + 0.004077159333974123, + 0.0036742109805345535, + 0.00360742025077343, + 0.010753038339316845, + 0.0034118916373699903, + 0.036480821669101715, + 0.031795550137758255, + 0.02751823328435421, + 0.04420514404773712, + 0.057036176323890686, + 0.03061646595597267, + 0.04722834378480911, + 0.02802063710987568, + 0.010393425822257996, + 0.013716400600969791, + 0.010041844099760056, + 0.012399469502270222, + 0.009916134178638458, + 0.008270639926195145, + 0.007555701769888401, + 0.00888271164149046, + 0.024977169930934906, + 0.025124911218881607, + 0.11961162090301514, + 0.012436548247933388, + 0.006419985089451075, + 0.01128559559583664, + 0.009359829127788544, + 0.10268339514732361, + 0.0022141262888908386, + 0.0017850922886282206, + 0.0024448013864457607, + 0.002631212119013071, + 0.0035302042961120605, + 0.002785195829346776, + 0.001733568962663412, + 0.002689522923901677, + 0.006854180246591568, + 0.004860257729887962, + 0.007379170972853899, + 0.004259204491972923, + 0.005352016072720289, + 0.004916350357234478, + 0.002793676918372512, + 0.005535346921533346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -1.5020420551300049 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_tensor([6276, 6276])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_tensor([6276, 6276])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..24102a82dd178dd6f95737619502a6f51fb398ac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_tensor([6276, 6276])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still close-up, focused horizontally on a mottled grotto wall. A warm, concentrated beam of light slants down from above and to the side, illuminating the details of the murals and casting delicate shadows across the uneven surface. At the center of the image is a vibrantly colored flying figure, draped in streamers of vermilion red and ochre green. She gracefully soars through the air, her face radiating a classical tranquility. Fluid lines define her voluptuous form, the rich texture of the mineral pigments and the subtle cracks of age clearly visible in the light. Then, this figure undergoes a miraculous transformation: the vibrant colors gradually lose saturation, and the image steadily and deliberately becomes transparent, as if a layer of ancient painting on glass is slowly dissolving. As her outline fades, the rough, ancient rock texture behind her gradually emerges, and sedimentary rock lines intertwined with dark brown and ochre penetrate from the flying fairy's body. Finally, the colorful fairy turns into an almost invisible ghostly afterimage, leaving only the original rock surface carrying thousands of years of history. The whole process is filled with the heavy sense of the passage of time and the sedimentation of history, creating a solemn and mysterious atmosphere.", + "seed": 23874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23874_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0627938974648714, + "vlm_logprob": -0.0627938974648714, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.0627938974648714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2114129215478897 + ], + "opposite_logprob_tokens": [ + -1.8364129066467285 + ], + "logprob_reward": -0.2114129215478897, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2114129215478897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3184206783771515, + 0.13222670555114746, + 0.10649152845144272, + 0.06145678833127022, + 0.051069241017103195, + 0.038346853107213974, + 0.1298397183418274, + 0.05448465421795845, + 0.05853625386953354, + 0.04912760853767395 + ], + "credit_map": [ + 0.027091605588793755, + 0.02184223383665085, + 0.07162067294120789, + 0.0250996146351099, + 0.03635149821639061, + 0.09820061922073364, + 0.01828943006694317, + 0.019924981519579887, + 0.0054878126829862595, + 0.018450036644935608, + 0.017604436725378036, + 0.015806354582309723, + 0.014557905495166779, + 0.014465371146798134, + 0.014146906323730946, + 0.031707875430583954, + 0.0058523803018033504, + 0.018448002636432648, + 0.006425598170608282, + 0.007311185356229544, + 0.031218774616718292, + 0.024396423250436783, + 0.00449071079492569, + 0.008348437957465649, + 0.007746262941509485, + 0.008118145167827606, + 0.006767216604202986, + 0.005364673212170601, + 0.012857457622885704, + 0.006244538817554712, + 0.007915441878139973, + 0.006443051155656576, + 0.0035080418456345797, + 0.004533218685537577, + 0.006618161220103502, + 0.006306047551333904, + 0.008580216206610203, + 0.008092698641121387, + 0.006394506897777319, + 0.007036344148218632, + 0.003358827205374837, + 0.003862321376800537, + 0.004802156705409288, + 0.006566161755472422, + 0.006606557872146368, + 0.003918133210390806, + 0.004808166995644569, + 0.004424525424838066, + 0.009514336474239826, + 0.013329783454537392, + 0.009381785057485104, + 0.013699542731046677, + 0.019020285457372665, + 0.015814023092389107, + 0.01780877448618412, + 0.031271178275346756, + 0.004590545315295458, + 0.0075712245889008045, + 0.005587202962487936, + 0.005583689082413912, + 0.005780672188848257, + 0.007700947113335133, + 0.005426737945526838, + 0.012243629433214664, + 0.003012078581377864, + 0.003614452900364995, + 0.004182104486972094, + 0.01753324083983898, + 0.004467146471142769, + 0.004623955115675926, + 0.0056806765496730804, + 0.015422593802213669, + 0.002908649854362011, + 0.005362953990697861, + 0.0036368349101394415, + 0.010653868317604065, + 0.00468762731179595, + 0.0068070488050580025, + 0.005998683162033558, + 0.009071934968233109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.2114129215478897 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804781526327133 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.009804781526327133, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804781526327133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5985589623451233, + 0.10964607447385788, + 0.04149593412876129, + 0.020593734458088875, + 0.052566517144441605, + 0.05408996343612671, + 0.0516367107629776, + 0.021014520898461342, + 0.029310520738363266, + 0.021087124943733215 + ], + "credit_map": [ + 0.037079762667417526, + 0.07623214274644852, + 0.06054472550749779, + 0.11816289275884628, + 0.06831986457109451, + 0.052448634058237076, + 0.11789859086275101, + 0.06787234544754028, + 0.006267784629017115, + 0.029817935079336166, + 0.008814175613224506, + 0.015383879654109478, + 0.020919321104884148, + 0.010621326044201851, + 0.010799020528793335, + 0.007022630423307419, + 0.0035527723375707865, + 0.005984018091112375, + 0.00333650060929358, + 0.005706343799829483, + 0.005964659620076418, + 0.006354727316647768, + 0.003546473802998662, + 0.007050436921417713, + 0.002865414833649993, + 0.0023363393265753984, + 0.0023132008500397205, + 0.002040240215137601, + 0.0031139785423874855, + 0.003061936004087329, + 0.0019662680570036173, + 0.0028963570948690176, + 0.006887633353471756, + 0.005812859162688255, + 0.005341806914657354, + 0.006078456528484821, + 0.00791826844215393, + 0.006371166557073593, + 0.008841856382787228, + 0.0053144702687859535, + 0.007371425628662109, + 0.009469221346080303, + 0.0051365382969379425, + 0.0055244858376681805, + 0.007105899043381214, + 0.004755116533488035, + 0.007984643802046776, + 0.006742632947862148, + 0.006386244669556618, + 0.009740659035742283, + 0.0054641179740428925, + 0.005216946359723806, + 0.01004130020737648, + 0.004413769114762545, + 0.0050747087225317955, + 0.005298965144902468, + 0.0016920772613957524, + 0.0027948052156716585, + 0.002186311176046729, + 0.003098434302955866, + 0.003316903254017234, + 0.0022680547554045916, + 0.0027686895336955786, + 0.002889245515689254, + 0.0023473170585930347, + 0.003552067093551159, + 0.0024777723010629416, + 0.005447301547974348, + 0.0027264233212918043, + 0.002541732043027878, + 0.0033089960925281048, + 0.0069089108146727085, + 0.0016651564510539174, + 0.0032633519731462, + 0.0023000652436167, + 0.003168112365528941, + 0.0022716126404702663, + 0.0032326309010386467, + 0.001947340089827776, + 0.003238854929804802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009804781526327133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015166220255196095 + ], + "opposite_logprob_tokens": [ + -6.501516819000244 + ], + "logprob_reward": -0.0015166220255196095, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015166220255196095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32219308614730835, + 0.11460064351558685, + 0.04908622056245804, + 0.0828840509057045, + 0.03000701405107975, + 0.11384465545415878, + 0.12696468830108643, + 0.0602138489484787, + 0.050548627972602844, + 0.04965713620185852 + ], + "credit_map": [ + 0.017668690532445908, + 0.043357450515031815, + 0.03405256196856499, + 0.052561499178409576, + 0.03964490070939064, + 0.05499045178294182, + 0.04873104393482208, + 0.03118647076189518, + 0.004887813702225685, + 0.022387510165572166, + 0.008983268402516842, + 0.011246011592447758, + 0.02979227341711521, + 0.01684611104428768, + 0.01274631917476654, + 0.007711328100413084, + 0.003752039512619376, + 0.006098555400967598, + 0.005817153956741095, + 0.009300162084400654, + 0.005251632072031498, + 0.00698446249589324, + 0.006241575814783573, + 0.0056406413204967976, + 0.006475232075899839, + 0.005899498704820871, + 0.006156033370643854, + 0.009067672304809093, + 0.008427329361438751, + 0.01755717024207115, + 0.011361687444150448, + 0.017939427867531776, + 0.0024988732766360044, + 0.00276623060926795, + 0.0036529991775751114, + 0.0030487182084470987, + 0.00519964937120676, + 0.00451152678579092, + 0.004179062787443399, + 0.004149952437728643, + 0.010826141573488712, + 0.015290106646716595, + 0.010890573263168335, + 0.01486281119287014, + 0.01762874610722065, + 0.01418444886803627, + 0.018859969452023506, + 0.011301860213279724, + 0.010778332129120827, + 0.013558918610215187, + 0.01268791127949953, + 0.01809980906546116, + 0.027049705386161804, + 0.009886244311928749, + 0.018145719543099403, + 0.016758037731051445, + 0.004270927514880896, + 0.006767627317458391, + 0.0068094441667199135, + 0.007371277082711458, + 0.007548986934125423, + 0.008451884612441063, + 0.009579319506883621, + 0.00941438041627407, + 0.0032915337942540646, + 0.004297618754208088, + 0.0033537792041897774, + 0.011885888874530792, + 0.003935346845537424, + 0.007314201444387436, + 0.005431585479527712, + 0.011038674972951412, + 0.0037147116381675005, + 0.007403088267892599, + 0.0058274297043681145, + 0.007859270088374615, + 0.006965636275708675, + 0.0065800161100924015, + 0.006706413812935352, + 0.004600568674504757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0015166220255196095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32499033212661743, + 0.057152990251779556, + 0.055590394884347916, + 0.07807131856679916, + 0.05571530759334564, + 0.086940236389637, + 0.08384622633457184, + 0.08308398723602295, + 0.09238577634096146, + 0.08222344517707825 + ], + "credit_map": [ + 0.01754474826157093, + 0.041532427072525024, + 0.0361853688955307, + 0.05215321481227875, + 0.0369398258626461, + 0.05118698254227638, + 0.053625404834747314, + 0.035822346806526184, + 0.003040326526388526, + 0.010146724060177803, + 0.0065694949589669704, + 0.0080524617806077, + 0.009523562155663967, + 0.006249212194234133, + 0.009043652564287186, + 0.004527552053332329, + 0.003924725577235222, + 0.00844119768589735, + 0.004344767890870571, + 0.005805082619190216, + 0.012064953334629536, + 0.00918674934655428, + 0.006277411710470915, + 0.005545508116483688, + 0.006591074634343386, + 0.006220766808837652, + 0.006000137887895107, + 0.007676159963011742, + 0.011440358124673367, + 0.014167173765599728, + 0.008312651887536049, + 0.017663000151515007, + 0.0066251386888325214, + 0.006160206161439419, + 0.006664695218205452, + 0.005667412653565407, + 0.01014416292309761, + 0.007367006037384272, + 0.007572578266263008, + 0.005514106713235378, + 0.008855057880282402, + 0.013105347752571106, + 0.00687363650649786, + 0.009541128762066364, + 0.012566003017127514, + 0.009864363819360733, + 0.018258346244692802, + 0.00787635799497366, + 0.005578347481787205, + 0.005922039970755577, + 0.008235935121774673, + 0.0133541664108634, + 0.02044621668756008, + 0.008281827904284, + 0.008818565867841244, + 0.013209125958383083, + 0.004402489401400089, + 0.010207351297140121, + 0.0071332394145429134, + 0.01659371517598629, + 0.009959833696484566, + 0.009565339423716068, + 0.012609266676008701, + 0.012612750753760338, + 0.004734344780445099, + 0.00898013450205326, + 0.007784262299537659, + 0.020122304558753967, + 0.009633274748921394, + 0.014328492805361748, + 0.010245424695312977, + 0.01655753329396248, + 0.005967351142317057, + 0.01385504100471735, + 0.009304264560341835, + 0.011601475067436695, + 0.011373181827366352, + 0.013289528898894787, + 0.00805989932268858, + 0.008772703818976879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.002195806708186865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903935551643372 + ], + "opposite_logprob_tokens": [ + -2.4640393257141113 + ], + "logprob_reward": -0.08903935551643372, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903935551643372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38288354873657227, + 0.04251798242330551, + 0.07753528654575348, + 0.03540165349841118, + 0.03821646794676781, + 0.11209998279809952, + 0.0983944907784462, + 0.06912096589803696, + 0.09273625165224075, + 0.05109332129359245 + ], + "credit_map": [ + 0.026029488071799278, + 0.03707173094153404, + 0.06188353896141052, + 0.049477871507406235, + 0.04359894618391991, + 0.09180613607168198, + 0.0417618565261364, + 0.03125396743416786, + 0.0036994507536292076, + 0.008464578539133072, + 0.005886595696210861, + 0.005810758098959923, + 0.005059470888227224, + 0.003520706668496132, + 0.005447831470519304, + 0.004628590773791075, + 0.005947122350335121, + 0.0164999607950449, + 0.006540482863783836, + 0.011911232955753803, + 0.01560355443507433, + 0.007681637071073055, + 0.006809349171817303, + 0.006541942711919546, + 0.003040269948542118, + 0.003904833924025297, + 0.0029792250134050846, + 0.003519974183291197, + 0.007849096320569515, + 0.004087834618985653, + 0.003802469465881586, + 0.00621794955804944, + 0.0037826818879693747, + 0.004794561769813299, + 0.0036993499379605055, + 0.004187153652310371, + 0.005685748066753149, + 0.006678441539406776, + 0.004816649481654167, + 0.00457187881693244, + 0.010936912149190903, + 0.015665601938962936, + 0.010160641744732857, + 0.016974560916423798, + 0.015761815011501312, + 0.014855688437819481, + 0.01893514394760132, + 0.008809614926576614, + 0.007105839438736439, + 0.008618428371846676, + 0.008360794745385647, + 0.009634347632527351, + 0.013541333377361298, + 0.011879471130669117, + 0.013791393488645554, + 0.025462878867983818, + 0.004247867036610842, + 0.007860120385885239, + 0.005301762372255325, + 0.013256681151688099, + 0.00834070984274149, + 0.012234888039529324, + 0.008469300344586372, + 0.009409631602466106, + 0.004170258063822985, + 0.005805456545203924, + 0.005233755800873041, + 0.025392018258571625, + 0.008123701438307762, + 0.011311055161058903, + 0.0089450404047966, + 0.023754963651299477, + 0.004721449688076973, + 0.00789837446063757, + 0.005587239749729633, + 0.007171251345425844, + 0.0064941816963255405, + 0.007298257201910019, + 0.0058927880600094795, + 0.006029775831848383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.08903935551643372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still close-up, focused horizontally on a mottled grotto wall. A warm, concentrated beam of light slants down from above and to the side, illuminating the details of the murals and casting delicate shadows across the uneven surface. At the center of the image is a vibrantly colored flying figure, draped in streamers of vermilion red and ochre green. She gracefully soars through the air, her face radiating a classical tranquility. Fluid lines define her voluptuous form, the rich texture of the mineral pigments and the subtle cracks of age clearly visible in the light. Then, this figure undergoes a miraculous transformation: the vibrant colors gradually lose saturation, and the image steadily and deliberately becomes transparent, as if a layer of ancient painting on glass is slowly dissolving. As her outline fades, the rough, ancient rock texture behind her gradually emerges, and sedimentary rock lines intertwined with dark brown and ochre penetrate from the flying fairy's body. Finally, the colorful fairy turns into an almost invisible ghostly afterimage, leaving only the original rock surface carrying thousands of years of history. The whole process is filled with the heavy sense of the passage of time and the sedimentation of history, creating a solemn and mysterious atmosphere.", + "seed": 23875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23875_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.055521893035620454, + "vlm_logprob": -0.055521893035620454, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.055521893035620454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1216372400522232 + ], + "opposite_logprob_tokens": [ + -2.4966373443603516 + ], + "logprob_reward": -0.1216372400522232, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1216372400522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1329270899295807, + 0.1477927565574646, + 0.16322985291481018, + 0.06604699045419693, + 0.19860561192035675, + 0.06771243363618851, + 0.03756925091147423, + 0.07407034188508987, + 0.0519668348133564, + 0.06007881462574005 + ], + "credit_map": [ + 0.015389380976557732, + 0.022429130971431732, + 0.020001497119665146, + 0.02035137265920639, + 0.01535620167851448, + 0.013757124543190002, + 0.012967094779014587, + 0.012675291858613491, + 0.012578058987855911, + 0.027083290740847588, + 0.014636187814176083, + 0.021608946844935417, + 0.021520383656024933, + 0.01593342237174511, + 0.01309347152709961, + 0.021338995546102524, + 0.010912907309830189, + 0.013823356479406357, + 0.011333029717206955, + 0.012657876126468182, + 0.011665326543152332, + 0.047082044184207916, + 0.010906202718615532, + 0.04484911635518074, + 0.0048513952642679214, + 0.008197122253477573, + 0.008764506317675114, + 0.004925561137497425, + 0.015050625428557396, + 0.005482105538249016, + 0.005760141648352146, + 0.013015532866120338, + 0.021534482017159462, + 0.021693209186196327, + 0.01408546231687069, + 0.02988593839108944, + 0.027586327865719795, + 0.029509281739592552, + 0.031095219776034355, + 0.02321568690240383, + 0.005825707223266363, + 0.008395608514547348, + 0.006906209979206324, + 0.010251158848404884, + 0.009911676868796349, + 0.007739738095551729, + 0.012175778858363628, + 0.006506561301648617, + 0.003649237332865596, + 0.0037385583855211735, + 0.004284117370843887, + 0.0049761696718633175, + 0.0039344788528978825, + 0.0061394209042191505, + 0.006422335747629404, + 0.0044249300844967365, + 0.004723615013062954, + 0.012537947855889797, + 0.006960427854210138, + 0.01885281316936016, + 0.008787441998720169, + 0.007079234346747398, + 0.009733118116855621, + 0.005395748186856508, + 0.006287712603807449, + 0.008225072175264359, + 0.007631469983607531, + 0.009342160075902939, + 0.004347899463027716, + 0.0037955669686198235, + 0.007985251955688, + 0.0043517048470675945, + 0.007208222057670355, + 0.009726828895509243, + 0.004705601371824741, + 0.0055293249897658825, + 0.007597801741212606, + 0.010465487837791443, + 0.0077169123105704784, + 0.007128634490072727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.1216372400522232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902072370052338 + ], + "opposite_logprob_tokens": [ + -2.5790207386016846 + ], + "logprob_reward": -0.07902072370052338, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902072370052338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1137174665927887, + 0.05898338556289673, + 0.11004558205604553, + 0.05639999732375145, + 0.40343326330184937, + 0.03894659876823425, + 0.05975095555186272, + 0.08064291626214981, + 0.04978837072849274, + 0.028291406109929085 + ], + "credit_map": [ + 0.013536103069782257, + 0.01690065488219261, + 0.017628641799092293, + 0.02062605507671833, + 0.01256977766752243, + 0.010016853921115398, + 0.012806317768990993, + 0.009633070789277554, + 0.0059715332463383675, + 0.008066779002547264, + 0.00426026014611125, + 0.005564795807003975, + 0.011052927933633327, + 0.007095692213624716, + 0.007838426157832146, + 0.009132969193160534, + 0.004670894239097834, + 0.011206663213670254, + 0.007463761139661074, + 0.004566472955048084, + 0.007757203653454781, + 0.03394237905740738, + 0.007047956343740225, + 0.03339025378227234, + 0.005237999837845564, + 0.008843234740197659, + 0.00897269044071436, + 0.0048203663900494576, + 0.008806110359728336, + 0.00590056087821722, + 0.007311535999178886, + 0.00650750333443284, + 0.03814735636115074, + 0.03683864325284958, + 0.02790205180644989, + 0.06279754638671875, + 0.0541573241353035, + 0.06149749085307121, + 0.07327347248792648, + 0.048819396644830704, + 0.0041358317248523235, + 0.003972281701862812, + 0.004152591805905104, + 0.00411582738161087, + 0.005629010498523712, + 0.006176136899739504, + 0.007603219710290432, + 0.003161701373755932, + 0.005682099144905806, + 0.005899262148886919, + 0.008787058293819427, + 0.00643895473331213, + 0.007029591128230095, + 0.008702225051820278, + 0.011616041883826256, + 0.00559572596102953, + 0.00533631443977356, + 0.01939837820827961, + 0.006603863090276718, + 0.022338541224598885, + 0.00815924908965826, + 0.005577513948082924, + 0.008050771430134773, + 0.005178282968699932, + 0.0027350077871233225, + 0.008635303936898708, + 0.004471305292099714, + 0.01685469225049019, + 0.004484912380576134, + 0.0025436407886445522, + 0.0074698906391859055, + 0.002593615557998419, + 0.002521116053685546, + 0.003929516300559044, + 0.002514456631615758, + 0.0033914693631231785, + 0.003143233945593238, + 0.005434883292764425, + 0.002989906817674637, + 0.004366824869066477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07902072370052338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101674538105726 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101674538105726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101674538105726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384645253419876, + 0.055637892335653305, + 0.06500179320573807, + 0.06537529081106186, + 0.10031309723854065, + 0.07813411206007004, + 0.04336582496762276, + 0.17648328840732574, + 0.1653282791376114, + 0.11189597100019455 + ], + "credit_map": [ + 0.015959670767188072, + 0.026357637718319893, + 0.010707076638936996, + 0.013247927650809288, + 0.02090987004339695, + 0.026459256187081337, + 0.015331553295254707, + 0.009491530247032642, + 0.004569155629724264, + 0.008316585794091225, + 0.004606273490935564, + 0.0065743508748710155, + 0.008710709400475025, + 0.010242163203656673, + 0.006298347841948271, + 0.0063203019089996815, + 0.008502022363245487, + 0.011411120183765888, + 0.009724157862365246, + 0.005641990806907415, + 0.006727138999849558, + 0.00784213561564684, + 0.006822766270488501, + 0.008330458775162697, + 0.0055484953336417675, + 0.012197144329547882, + 0.008311675861477852, + 0.007649956736713648, + 0.007393519394099712, + 0.008135298267006874, + 0.00976579636335373, + 0.0063734096474945545, + 0.013358939439058304, + 0.010561499744653702, + 0.012638305313885212, + 0.016869084909558296, + 0.013458957895636559, + 0.011034755036234856, + 0.01322743110358715, + 0.009164116345345974, + 0.008360369130969048, + 0.010158081538975239, + 0.007461042609065771, + 0.013471855781972408, + 0.010795150883495808, + 0.009885366074740887, + 0.008195671252906322, + 0.009806569665670395, + 0.0032844175584614277, + 0.003896808484569192, + 0.0059625376015901566, + 0.00644234474748373, + 0.00447611790150404, + 0.005923881661146879, + 0.00948314554989338, + 0.0038965721614658833, + 0.010477223433554173, + 0.04586201161146164, + 0.013437536545097828, + 0.052868910133838654, + 0.019635619595646858, + 0.01103181578218937, + 0.015213510021567345, + 0.007956656627357006, + 0.010136798955500126, + 0.03033905290067196, + 0.014738180674612522, + 0.05860300734639168, + 0.01471265684813261, + 0.006183928810060024, + 0.024266652762889862, + 0.006347992457449436, + 0.012078662402927876, + 0.014821787364780903, + 0.010169734247028828, + 0.01327119767665863, + 0.014207459054887295, + 0.02178863063454628, + 0.011488069780170918, + 0.014070430770516396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.004101674538105726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818899229168892 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.002818899229168892, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818899229168892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2080688327550888, + 0.08171498775482178, + 0.09297861903905869, + 0.059776052832603455, + 0.08984537422657013, + 0.05642033740878105, + 0.04374125227332115, + 0.2288026511669159, + 0.0617973655462265, + 0.07685448974370956 + ], + "credit_map": [ + 0.021016910672187805, + 0.035703763365745544, + 0.021445654332637787, + 0.019701184704899788, + 0.02947785146534443, + 0.043783679604530334, + 0.0220718402415514, + 0.014867949299514294, + 0.008910813368856907, + 0.014324923045933247, + 0.007591633126139641, + 0.00931267999112606, + 0.010207859799265862, + 0.009474512189626694, + 0.009830744005739689, + 0.012061833404004574, + 0.008532732725143433, + 0.013284597545862198, + 0.008962241001427174, + 0.0063124471344053745, + 0.0077917613089084625, + 0.022349214181303978, + 0.008670483715832233, + 0.01707514375448227, + 0.004417468328028917, + 0.010391389951109886, + 0.008255028165876865, + 0.005730084143579006, + 0.011670448817312717, + 0.005921920295804739, + 0.00673675024881959, + 0.006652961019426584, + 0.010338540188968182, + 0.01168824452906847, + 0.013310080394148827, + 0.016054030507802963, + 0.01143356878310442, + 0.007921192795038223, + 0.008951855823397636, + 0.010147861205041409, + 0.007575212512165308, + 0.006912484299391508, + 0.005340029019862413, + 0.0066599673591554165, + 0.009426512755453587, + 0.006521460134536028, + 0.007990884594619274, + 0.005993792321532965, + 0.003711152821779251, + 0.004754099994897842, + 0.005024588666856289, + 0.0053994921036064625, + 0.004794473759829998, + 0.00657866382971406, + 0.009459133259952068, + 0.0040196506306529045, + 0.01450590044260025, + 0.05418211594223976, + 0.019102957099676132, + 0.06654679030179977, + 0.02882804162800312, + 0.013260272331535816, + 0.022128401324152946, + 0.010248200036585331, + 0.011613273061811924, + 0.008791188709437847, + 0.006418026518076658, + 0.010382852517068386, + 0.005718549247831106, + 0.004663852043449879, + 0.009123831056058407, + 0.005085795186460018, + 0.006645822897553444, + 0.007650235202163458, + 0.00623788358643651, + 0.012419900856912136, + 0.00871789176017046, + 0.01415849570184946, + 0.00857258215546608, + 0.012451685965061188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.002818899229168892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003092765808105 + ], + "opposite_logprob_tokens": [ + -2.695030927658081 + ], + "logprob_reward": -0.07003092765808105, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003092765808105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14656956493854523, + 0.11686072498559952, + 0.04606470465660095, + 0.031279295682907104, + 0.05923145264387131, + 0.030997199937701225, + 0.023105792701244354, + 0.32795166969299316, + 0.1700875163078308, + 0.04785211384296417 + ], + "credit_map": [ + 0.01738634891808033, + 0.02521318569779396, + 0.015922248363494873, + 0.02070598118007183, + 0.017972858622670174, + 0.017430376261472702, + 0.017095966264605522, + 0.014842609874904156, + 0.009832172654569149, + 0.02458660677075386, + 0.0082540987059474, + 0.015234052203595638, + 0.01606929302215576, + 0.015371017158031464, + 0.0108305299654603, + 0.0166829451918602, + 0.006052323151379824, + 0.005928550381213427, + 0.008238162845373154, + 0.005165214650332928, + 0.005519305355846882, + 0.004517297260463238, + 0.005925236735492945, + 0.004718618001788855, + 0.002505493350327015, + 0.004658524412661791, + 0.00458399998024106, + 0.0027858521789312363, + 0.004788659978657961, + 0.004005792085081339, + 0.004151772707700729, + 0.0037992007564753294, + 0.008823433890938759, + 0.005648569669574499, + 0.01537244115024805, + 0.006501420401036739, + 0.0049103545024991035, + 0.005114774219691753, + 0.005034327041357756, + 0.007826128043234348, + 0.0031326517928391695, + 0.0035300925374031067, + 0.0031208284199237823, + 0.003641572082415223, + 0.005201756488531828, + 0.004554579965770245, + 0.0031402118038386106, + 0.004675506614148617, + 0.002443520352244377, + 0.0021745713893324137, + 0.002852000528946519, + 0.0025921280030161142, + 0.0030451230704784393, + 0.0029977536760270596, + 0.0050006224773824215, + 0.0020000736694782972, + 0.02051086351275444, + 0.061302538961172104, + 0.028189007192850113, + 0.08286967873573303, + 0.056699950248003006, + 0.02135280892252922, + 0.03770726919174194, + 0.019319547340273857, + 0.011254334822297096, + 0.026469755917787552, + 0.014502612873911858, + 0.05585133656859398, + 0.015942098572850227, + 0.009755539707839489, + 0.02727903611958027, + 0.00903278961777687, + 0.006273652892559767, + 0.007909908890724182, + 0.0031804442405700684, + 0.004365589004009962, + 0.004964710213243961, + 0.008234853856265545, + 0.005662444047629833, + 0.007260510232299566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07003092765808105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still close-up, focused horizontally on a mottled grotto wall. A warm, concentrated beam of light slants down from above and to the side, illuminating the details of the murals and casting delicate shadows across the uneven surface. At the center of the image is a vibrantly colored flying figure, draped in streamers of vermilion red and ochre green. She gracefully soars through the air, her face radiating a classical tranquility. Fluid lines define her voluptuous form, the rich texture of the mineral pigments and the subtle cracks of age clearly visible in the light. Then, this figure undergoes a miraculous transformation: the vibrant colors gradually lose saturation, and the image steadily and deliberately becomes transparent, as if a layer of ancient painting on glass is slowly dissolving. As her outline fades, the rough, ancient rock texture behind her gradually emerges, and sedimentary rock lines intertwined with dark brown and ochre penetrate from the flying fairy's body. Finally, the colorful fairy turns into an almost invisible ghostly afterimage, leaving only the original rock surface carrying thousands of years of history. The whole process is filled with the heavy sense of the passage of time and the sedimentation of history, creating a solemn and mysterious atmosphere.", + "seed": 23876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23876_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06968640852719546, + "vlm_logprob": -0.06968640852719546, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.06968640852719546, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24700085818767548 + ], + "opposite_logprob_tokens": [ + -1.622000813484192 + ], + "logprob_reward": -0.24700085818767548, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24700085818767548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11791834235191345, + 0.05264023318886757, + 0.0792456716299057, + 0.06719537824392319, + 0.11152126640081406, + 0.15287281572818756, + 0.09503687918186188, + 0.15690438449382782, + 0.10023705661296844, + 0.066428042948246 + ], + "credit_map": [ + 0.015407497994601727, + 0.014084680937230587, + 0.018910542130470276, + 0.019231457263231277, + 0.013105452060699463, + 0.01205659843981266, + 0.010273138992488384, + 0.014848961494863033, + 0.004551148973405361, + 0.004133255686610937, + 0.007576554547995329, + 0.011276882141828537, + 0.005615733563899994, + 0.00891287624835968, + 0.004339341074228287, + 0.006234439089894295, + 0.011724768206477165, + 0.006221380550414324, + 0.014583248645067215, + 0.011754998937249184, + 0.005250594578683376, + 0.008542854338884354, + 0.008741279132664204, + 0.01242653839290142, + 0.006607663352042437, + 0.0052644358947873116, + 0.013820032589137554, + 0.009961150586605072, + 0.004610163159668446, + 0.00540895713493228, + 0.008328202180564404, + 0.013194768689572811, + 0.01043175533413887, + 0.014931670390069485, + 0.02396538108587265, + 0.017157159745693207, + 0.01130782999098301, + 0.010679625906050205, + 0.010698383674025536, + 0.012349451892077923, + 0.00870432984083891, + 0.008356205187737942, + 0.012818423099815845, + 0.05218985304236412, + 0.012264843098819256, + 0.011534889228641987, + 0.0173321682959795, + 0.029672089964151382, + 0.007508800830692053, + 0.013050195761024952, + 0.008193749003112316, + 0.011837484315037727, + 0.014964940957725048, + 0.011762972921133041, + 0.016435252502560616, + 0.011283465661108494, + 0.01936613954603672, + 0.015409290790557861, + 0.01977953314781189, + 0.02318616211414337, + 0.02332914248108864, + 0.017068961635231972, + 0.023035045713186264, + 0.015730084851384163, + 0.030749181285500526, + 0.008488439954817295, + 0.009375089779496193, + 0.010462120175361633, + 0.010203039273619652, + 0.010023827664554119, + 0.010798262432217598, + 0.010137083940207958, + 0.013221126981079578, + 0.005387707147747278, + 0.007282068021595478, + 0.008194905705749989, + 0.007859615609049797, + 0.010003846138715744, + 0.008524556644260883, + 0.00595420878380537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.24700085818767548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005281068850308657 + ], + "opposite_logprob_tokens": [ + -5.2552809715271 + ], + "logprob_reward": -0.005281068850308657, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005281068850308657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13011065125465393, + 0.10652124881744385, + 0.10833068937063217, + 0.11238711327314377, + 0.14719313383102417, + 0.15306225419044495, + 0.04722185060381889, + 0.03646509721875191, + 0.09211387485265732, + 0.06659411638975143 + ], + "credit_map": [ + 0.01159516628831625, + 0.014147639274597168, + 0.019740965217351913, + 0.01923518255352974, + 0.010908945463597775, + 0.020095329731702805, + 0.013601378537714481, + 0.020786048844456673, + 0.007995730265974998, + 0.009253880009055138, + 0.01482014637440443, + 0.02461581863462925, + 0.009280722588300705, + 0.01775352656841278, + 0.010039745829999447, + 0.012761672958731651, + 0.013222389854490757, + 0.010162187740206718, + 0.0148369912058115, + 0.013029279187321663, + 0.019104698672890663, + 0.013595884665846825, + 0.010712709277868271, + 0.013666546903550625, + 0.008409356698393822, + 0.009027784690260887, + 0.027501503005623817, + 0.015579838305711746, + 0.00949198380112648, + 0.010037314146757126, + 0.009855077601969242, + 0.022484252229332924, + 0.012351096607744694, + 0.013314501382410526, + 0.020208485424518585, + 0.03655872493982315, + 0.013976510614156723, + 0.01646498218178749, + 0.01860509067773819, + 0.01571374386548996, + 0.011333879083395004, + 0.01282212883234024, + 0.011932509019970894, + 0.05049923062324524, + 0.011751921847462654, + 0.014320509508252144, + 0.010991310700774193, + 0.029410744085907936, + 0.0038088031578809023, + 0.004341984633356333, + 0.004950581584125757, + 0.006886782590299845, + 0.007149760611355305, + 0.006734278053045273, + 0.008633674122393131, + 0.004715983755886555, + 0.0068868910893797874, + 0.004228788893669844, + 0.003317486261948943, + 0.003729608142748475, + 0.004524208139628172, + 0.004396381787955761, + 0.00459252018481493, + 0.004789212718605995, + 0.027372024953365326, + 0.008942211046814919, + 0.009002539329230785, + 0.009046330116689205, + 0.008210535161197186, + 0.011023317463696003, + 0.009020023979246616, + 0.009496886283159256, + 0.020567679777741432, + 0.00721956230700016, + 0.006607365794479847, + 0.005505249835550785, + 0.006284619215875864, + 0.007529543247073889, + 0.006304045207798481, + 0.0065760537981987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.005281068850308657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061344001442193985, + 0.021931249648332596, + 0.019040662795305252, + 0.01854337938129902, + 0.06250030547380447, + 0.44552144408226013, + 0.10601668804883957, + 0.0677730068564415, + 0.14018544554710388, + 0.05714389681816101 + ], + "credit_map": [ + 0.006918004713952541, + 0.005761579144746065, + 0.010008621029555798, + 0.008137225173413754, + 0.007530753966420889, + 0.007763882167637348, + 0.006719887722283602, + 0.008504047989845276, + 0.002734364243224263, + 0.0017094248905777931, + 0.0031329388730227947, + 0.003479697974398732, + 0.002850903896614909, + 0.002508755773305893, + 0.0023606433533132076, + 0.00315452110953629, + 0.002720564603805542, + 0.001499927369877696, + 0.0026323527563363314, + 0.0027496202383190393, + 0.00213870732113719, + 0.0025616162456572056, + 0.002021665219217539, + 0.0027162095066159964, + 0.0026588698383420706, + 0.0022083644289523363, + 0.0028431653045117855, + 0.0024496749974787235, + 0.002111167646944523, + 0.0016255279770120978, + 0.0018508249195292592, + 0.0027957854326814413, + 0.005573219154030085, + 0.004672338720411062, + 0.01256585493683815, + 0.01377870887517929, + 0.007810440380126238, + 0.005649187136441469, + 0.006389703135937452, + 0.006060851737856865, + 0.024659784510731697, + 0.022674934938549995, + 0.035818349570035934, + 0.1510351151227951, + 0.03711126372218132, + 0.04265213385224342, + 0.047017861157655716, + 0.08455199748277664, + 0.00838373601436615, + 0.011006500571966171, + 0.011740879155695438, + 0.012871275655925274, + 0.014123553410172462, + 0.01486477255821228, + 0.022295471280813217, + 0.010730501264333725, + 0.01541202887892723, + 0.007347159553319216, + 0.008482187055051327, + 0.008501083590090275, + 0.0064900158904492855, + 0.008108348585665226, + 0.006814912892878056, + 0.006617262028157711, + 0.041736237704753876, + 0.012711672112345695, + 0.011687411926686764, + 0.013279170729219913, + 0.01133321039378643, + 0.018258297815918922, + 0.014985601417720318, + 0.016193853691220284, + 0.01639232411980629, + 0.006109970156103373, + 0.005498483311384916, + 0.00434159766882658, + 0.00510625122115016, + 0.008043655194342136, + 0.005875898990780115, + 0.005775711499154568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634591285139322 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634591285139322, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634591285139322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10825368762016296, + 0.02776859886944294, + 0.02209986187517643, + 0.025180144235491753, + 0.053200192749500275, + 0.21481841802597046, + 0.1321457177400589, + 0.12430872023105621, + 0.17345139384269714, + 0.11877328157424927 + ], + "credit_map": [ + 0.01240474358201027, + 0.009885469451546669, + 0.014884615316987038, + 0.01801387593150139, + 0.013737654313445091, + 0.010942383669316769, + 0.010934941470623016, + 0.017449993640184402, + 0.0028442617040127516, + 0.0020407543051987886, + 0.004030459560453892, + 0.0041534509509801865, + 0.0029849428683519363, + 0.004607253707945347, + 0.0033245068043470383, + 0.003782968269661069, + 0.002590365009382367, + 0.002373458119109273, + 0.00313195982016623, + 0.003531397320330143, + 0.002328039612621069, + 0.0030562833417207003, + 0.0026915939524769783, + 0.002396762603893876, + 0.002720862627029419, + 0.0015537554863840342, + 0.005730312317609787, + 0.002921612001955509, + 0.0022395129781216383, + 0.0020925370045006275, + 0.002336089266464114, + 0.005585461389273405, + 0.004661794286221266, + 0.004957179538905621, + 0.009665659628808498, + 0.007293430622667074, + 0.006648637820035219, + 0.005061909556388855, + 0.006114191375672817, + 0.008797388523817062, + 0.012547893449664116, + 0.0103981401771307, + 0.01860852725803852, + 0.06737975776195526, + 0.020099081099033356, + 0.02228056825697422, + 0.02798089198768139, + 0.03552353009581566, + 0.010832218453288078, + 0.01813119649887085, + 0.010626702569425106, + 0.016326742246747017, + 0.023010630160570145, + 0.016364436596632004, + 0.020802712067961693, + 0.016051068902015686, + 0.01863253116607666, + 0.011602952145040035, + 0.01648012176156044, + 0.019436953589320183, + 0.019134311005473137, + 0.0127433892339468, + 0.01439862884581089, + 0.011879827827215195, + 0.04421684518456459, + 0.015025660395622253, + 0.015114239417016506, + 0.021777989342808723, + 0.0186744537204504, + 0.021960003301501274, + 0.01837354712188244, + 0.018308648839592934, + 0.0196829866617918, + 0.012668195180594921, + 0.011009358800947666, + 0.015433340333402157, + 0.015379571355879307, + 0.018352273851633072, + 0.014441469684243202, + 0.011806081049144268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.004634591285139322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902310580015182 + ], + "opposite_logprob_tokens": [ + -2.4640231132507324 + ], + "logprob_reward": -0.08902310580015182, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902310580015182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.206289142370224, + 0.04961134493350983, + 0.02486545965075493, + 0.03995358943939209, + 0.026903994381427765, + 0.2226714938879013, + 0.06105886772274971, + 0.0699794813990593, + 0.12862315773963928, + 0.17004349827766418 + ], + "credit_map": [ + 0.019808104261755943, + 0.02858584001660347, + 0.026712844148278236, + 0.04137561842799187, + 0.022573966532945633, + 0.023163577541708946, + 0.020072709769010544, + 0.023996474221348763, + 0.005558207631111145, + 0.00578848272562027, + 0.005648894235491753, + 0.010405618697404861, + 0.004871627315878868, + 0.0073265815153717995, + 0.004385080188512802, + 0.005626850761473179, + 0.00294028059579432, + 0.0029467062558978796, + 0.0027266659308224916, + 0.005471229553222656, + 0.0016903856303542852, + 0.0024577646981924772, + 0.0035152563359588385, + 0.003117169253528118, + 0.004145453684031963, + 0.004144849721342325, + 0.005995933897793293, + 0.007165119051933289, + 0.004755531437695026, + 0.004369037691503763, + 0.004745580721646547, + 0.004632078111171722, + 0.0034461854957044125, + 0.0030974880792200565, + 0.0036663084756582975, + 0.0032464698888361454, + 0.0029997252859175205, + 0.0032413494773209095, + 0.003575907787308097, + 0.0036305582616478205, + 0.014110423624515533, + 0.015834368765354156, + 0.018804728984832764, + 0.06664379686117172, + 0.018037889152765274, + 0.02064027450978756, + 0.028375444933772087, + 0.040224555879831314, + 0.006875169929116964, + 0.012660326436161995, + 0.005702273920178413, + 0.007523423060774803, + 0.005635309033095837, + 0.0072386618703603745, + 0.007954003289341927, + 0.007469698786735535, + 0.009482312016189098, + 0.012117940932512283, + 0.0086516747251153, + 0.00968741625547409, + 0.007158505264669657, + 0.006933628115803003, + 0.009233278222382069, + 0.006714720278978348, + 0.040171921253204346, + 0.012858274392783642, + 0.010728906840085983, + 0.01075783185660839, + 0.009068713523447514, + 0.01600778102874756, + 0.013543931767344475, + 0.015485790558159351, + 0.050107281655073166, + 0.018253330141305923, + 0.014024997130036354, + 0.01308476086705923, + 0.01142994686961174, + 0.024819783866405487, + 0.018394362181425095, + 0.01992902345955372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.08902310580015182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still close-up, focused horizontally on a mottled grotto wall. A warm, concentrated beam of light slants down from above and to the side, illuminating the details of the murals and casting delicate shadows across the uneven surface. At the center of the image is a vibrantly colored flying figure, draped in streamers of vermilion red and ochre green. She gracefully soars through the air, her face radiating a classical tranquility. Fluid lines define her voluptuous form, the rich texture of the mineral pigments and the subtle cracks of age clearly visible in the light. Then, this figure undergoes a miraculous transformation: the vibrant colors gradually lose saturation, and the image steadily and deliberately becomes transparent, as if a layer of ancient painting on glass is slowly dissolving. As her outline fades, the rough, ancient rock texture behind her gradually emerges, and sedimentary rock lines intertwined with dark brown and ochre penetrate from the flying fairy's body. Finally, the colorful fairy turns into an almost invisible ghostly afterimage, leaving only the original rock surface carrying thousands of years of history. The whole process is filled with the heavy sense of the passage of time and the sedimentation of history, creating a solemn and mysterious atmosphere.", + "seed": 23877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_6/index_6276_seed_23877_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052526975004002455, + "vlm_logprob": -0.052526975004002455, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.052526975004002455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16336743533611298 + ], + "opposite_logprob_tokens": [ + -2.163367509841919 + ], + "logprob_reward": -0.16336743533611298, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.16336743533611298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.173323392868042, + 0.14759671688079834, + 0.06418225169181824, + 0.10901372134685516, + 0.19624881446361542, + 0.10274849832057953, + 0.03870362415909767, + 0.0403137281537056, + 0.0639185830950737, + 0.06395065784454346 + ], + "credit_map": [ + 0.010207741521298885, + 0.02758128196001053, + 0.010258421301841736, + 0.012358255684375763, + 0.04284931346774101, + 0.0408012755215168, + 0.01640786975622177, + 0.012859228067100048, + 0.010273096151649952, + 0.01554990652948618, + 0.018781457096338272, + 0.012636145576834679, + 0.026586472988128662, + 0.03453277423977852, + 0.012687867507338524, + 0.01654900051653385, + 0.004985360894352198, + 0.008137233555316925, + 0.006085922010242939, + 0.01472541969269514, + 0.008676248602569103, + 0.010168291628360748, + 0.007717642467468977, + 0.003686131676658988, + 0.009940425865352154, + 0.016782380640506744, + 0.013897493481636047, + 0.01709328405559063, + 0.01665838249027729, + 0.014634668827056885, + 0.011295244097709656, + 0.008711840957403183, + 0.012937488034367561, + 0.023789115250110626, + 0.022277798503637314, + 0.024205850437283516, + 0.02498520351946354, + 0.029930660501122475, + 0.024952728301286697, + 0.03316996619105339, + 0.006433340720832348, + 0.0067580947652459145, + 0.009159124456346035, + 0.008728815242648125, + 0.009320289827883244, + 0.02938694693148136, + 0.01763305813074112, + 0.015328836627304554, + 0.0034410201478749514, + 0.004540827125310898, + 0.0028679652605205774, + 0.0024235511664301157, + 0.006237589288502932, + 0.010000484064221382, + 0.003621206618845463, + 0.005570981185883284, + 0.004175445530563593, + 0.005218713544309139, + 0.008513834327459335, + 0.005054622422903776, + 0.007884384132921696, + 0.0028557039331644773, + 0.003769408445805311, + 0.0028416176792234182, + 0.006703674793243408, + 0.008032283745706081, + 0.006479436997324228, + 0.012605153024196625, + 0.00869416631758213, + 0.005571398418396711, + 0.010848600417375565, + 0.004983873106539249, + 0.008625604212284088, + 0.0091612683609128, + 0.007839995436370373, + 0.007697900757193565, + 0.010242978110909462, + 0.00563069200143218, + 0.009261000901460648, + 0.00549121480435133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.16336743533611298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096051782369614 + ], + "opposite_logprob_tokens": [ + -4.141096115112305 + ], + "logprob_reward": -0.016096051782369614, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096051782369614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17985956370830536, + 0.07229132950305939, + 0.04496463015675545, + 0.04770822077989578, + 0.029134826734662056, + 0.34981605410575867, + 0.20417629182338715, + 0.019971385598182678, + 0.027550917118787766, + 0.024526730179786682 + ], + "credit_map": [ + 0.01105944998562336, + 0.016524264588952065, + 0.007704901974648237, + 0.012358932755887508, + 0.0642469972372055, + 0.04582420736551285, + 0.013095217756927013, + 0.009045596234500408, + 0.00454670749604702, + 0.006473087705671787, + 0.012382824905216694, + 0.009967176243662834, + 0.010695328004658222, + 0.013765554875135422, + 0.0068267229944467545, + 0.007633927743881941, + 0.004157267045229673, + 0.006730364635586739, + 0.0065866210497915745, + 0.00639300188049674, + 0.005307537969201803, + 0.007300271652638912, + 0.0053109037689864635, + 0.0031786595936864614, + 0.003453122219070792, + 0.005294827278703451, + 0.013638289645314217, + 0.011667531915009022, + 0.003926817327737808, + 0.0029443034436553717, + 0.003448713105171919, + 0.0033346216659992933, + 0.0020681379828602076, + 0.005198576487600803, + 0.0027813510969281197, + 0.003490758826956153, + 0.005105701275169849, + 0.003588701831176877, + 0.0037320663686841726, + 0.0031695321667939425, + 0.04198131337761879, + 0.020099176093935966, + 0.02185720019042492, + 0.012161975726485252, + 0.12884080410003662, + 0.04380904138088226, + 0.06262864172458649, + 0.018437914550304413, + 0.02049584686756134, + 0.021013863384723663, + 0.012643644586205482, + 0.009119323454797268, + 0.032279640436172485, + 0.06210015341639519, + 0.012473403476178646, + 0.03405042365193367, + 0.0018825704464688897, + 0.0018163243075832725, + 0.003131372621282935, + 0.0021603000350296497, + 0.003738139057531953, + 0.002143188612535596, + 0.0026655085384845734, + 0.002433984773233533, + 0.003067696699872613, + 0.004245140589773655, + 0.0035879912320524454, + 0.00519300764426589, + 0.0030381185933947563, + 0.002500606467947364, + 0.0036205716896802187, + 0.002297787694260478, + 0.0028132530860602856, + 0.0028191397432237864, + 0.0030627045780420303, + 0.0032041750382632017, + 0.004204859025776386, + 0.0023565995506942272, + 0.003972052130848169, + 0.002093949355185032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.016096051782369614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017170461360365152 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017170461360365152, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017170461360365152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3312869369983673, + 0.03321778401732445, + 0.031450122594833374, + 0.029396532103419304, + 0.03637627139687538, + 0.12153299897909164, + 0.11895453929901123, + 0.17962415516376495, + 0.05517834797501564, + 0.06298238039016724 + ], + "credit_map": [ + 0.018548136577010155, + 0.02829030342400074, + 0.008862891234457493, + 0.018029719591140747, + 0.12659133970737457, + 0.10356111824512482, + 0.016546063125133514, + 0.010857349261641502, + 0.0029201162979006767, + 0.0029534490313380957, + 0.004927156027406454, + 0.00513882003724575, + 0.006678256206214428, + 0.004501869436353445, + 0.0032422798685729504, + 0.002855835948139429, + 0.003881502663716674, + 0.003811837872490287, + 0.0029449148569256067, + 0.005276657175272703, + 0.004151632077991962, + 0.004849401768296957, + 0.0035514701157808304, + 0.0029827046673744917, + 0.0024635749869048595, + 0.004146856255829334, + 0.006706546526402235, + 0.0027300638612359762, + 0.003032967448234558, + 0.0030069821514189243, + 0.0027072527445852757, + 0.004602285102009773, + 0.002673402428627014, + 0.003968133125454187, + 0.004872343968600035, + 0.0029570404440164566, + 0.005293746944516897, + 0.006255271378904581, + 0.0029862066730856895, + 0.007370122242718935, + 0.012264909222722054, + 0.010033763945102692, + 0.013868081383407116, + 0.006407937500625849, + 0.020608024671673775, + 0.030936500057578087, + 0.01090299803763628, + 0.01651078090071678, + 0.013650848530232906, + 0.007752653211355209, + 0.007267714943736792, + 0.005114508792757988, + 0.01472593005746603, + 0.038867779076099396, + 0.009103477001190186, + 0.022471623495221138, + 0.017423607409000397, + 0.018064526841044426, + 0.02748933807015419, + 0.016300778836011887, + 0.04844304546713829, + 0.01633664034307003, + 0.020546618849039078, + 0.015019586309790611, + 0.0070577748119831085, + 0.0073853954672813416, + 0.0061689214780926704, + 0.007799800019711256, + 0.008991971611976624, + 0.00435663154348731, + 0.008097267709672451, + 0.005320582073181868, + 0.0068087452091276646, + 0.006801186595112085, + 0.008009636774659157, + 0.008486486040055752, + 0.008555938489735126, + 0.005954709369689226, + 0.011168652214109898, + 0.007197016850113869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0017170461360365152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18898481130599976, + 0.030069593340158463, + 0.03788856416940689, + 0.0613868311047554, + 0.05220417305827141, + 0.3017631769180298, + 0.03413819149136543, + 0.12023712694644928, + 0.05817357823252678, + 0.11515398323535919 + ], + "credit_map": [ + 0.01099249254912138, + 0.01763448305428028, + 0.006702718324959278, + 0.011358162388205528, + 0.06809716671705246, + 0.05791551247239113, + 0.008584249764680862, + 0.007700016256421804, + 0.003567248350009322, + 0.003474024124443531, + 0.006045910529792309, + 0.003823550185188651, + 0.0032377743627876043, + 0.0031272172927856445, + 0.00416553346440196, + 0.0026283336337655783, + 0.00354577274993062, + 0.0035243425518274307, + 0.005140720400959253, + 0.004089106805622578, + 0.006146311294287443, + 0.0073461071588099, + 0.004703196696937084, + 0.003393002785742283, + 0.003751695156097412, + 0.006133836228400469, + 0.009097038768231869, + 0.009460638277232647, + 0.01059674471616745, + 0.009268845431506634, + 0.0082278186455369, + 0.004850203637033701, + 0.0032470962032675743, + 0.005080693867057562, + 0.004417452961206436, + 0.003958864603191614, + 0.006082627456635237, + 0.01269995141774416, + 0.004740262404084206, + 0.011977220885455608, + 0.027193203568458557, + 0.016282880678772926, + 0.015405922196805477, + 0.016314543783664703, + 0.11018320173025131, + 0.06172342598438263, + 0.032587938010692596, + 0.022072041407227516, + 0.0042356220073997974, + 0.0053846575319767, + 0.003558781463652849, + 0.0026471195742487907, + 0.006386762484908104, + 0.003514666575938463, + 0.005449183750897646, + 0.002961396239697933, + 0.012790783308446407, + 0.014566775411367416, + 0.01747053675353527, + 0.01208039466291666, + 0.025096632540225983, + 0.01082795113325119, + 0.01602078042924404, + 0.011383263394236565, + 0.006899270694702864, + 0.0076933265663683414, + 0.008854166604578495, + 0.009842891246080399, + 0.008516994304955006, + 0.004577734041959047, + 0.007127849385142326, + 0.004661343526095152, + 0.011099454015493393, + 0.012225200422108173, + 0.014916276559233665, + 0.016396034508943558, + 0.015187330543994904, + 0.010929769836366177, + 0.022305263206362724, + 0.012094643898308277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0024867109023034573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up focused horizontally on a mottled grotto wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vibrantly colored flying figure draped in vermilion red and ochre green streamers visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the figure undergo a transformation where colors lose saturation and the image becomes transparent to reveal the rock texture behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure eventually turn into an almost invisible ghostly afterimage leaving only the rock surface visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall atmosphere solemn and mysterious, conveying a heavy sense of the passage of time?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789676308631897 + ], + "opposite_logprob_tokens": [ + -2.578967571258545 + ], + "logprob_reward": -0.0789676308631897, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789676308631897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18378494679927826, + 0.07047665119171143, + 0.044002510607242584, + 0.056614767760038376, + 0.05519590899348259, + 0.22884881496429443, + 0.12630441784858704, + 0.062127675861120224, + 0.054127488285303116, + 0.11851685494184494 + ], + "credit_map": [ + 0.008116328157484531, + 0.015103175304830074, + 0.005895256996154785, + 0.01085353922098875, + 0.06805770099163055, + 0.059330832213163376, + 0.009222017601132393, + 0.007206094916909933, + 0.00451029883697629, + 0.004635591059923172, + 0.013246909715235233, + 0.005842790938913822, + 0.014493852853775024, + 0.012414432130753994, + 0.006605743896216154, + 0.008727036416530609, + 0.0040276506915688515, + 0.007848062552511692, + 0.00846182368695736, + 0.006531912833452225, + 0.005904121790081263, + 0.0040467833168804646, + 0.0036005170550197363, + 0.0035816412419080734, + 0.004154285416007042, + 0.007077122572809458, + 0.005167498253285885, + 0.009279404766857624, + 0.010305896401405334, + 0.007980424910783768, + 0.008039161562919617, + 0.0046109771355986595, + 0.003235855372622609, + 0.005979056004434824, + 0.005385234486311674, + 0.005328131839632988, + 0.007023341488093138, + 0.008511845022439957, + 0.009188654832541943, + 0.01054378878325224, + 0.01116230059415102, + 0.015839174389839172, + 0.012016097083687782, + 0.010760855861008167, + 0.08100532740354538, + 0.053143586963415146, + 0.02107158862054348, + 0.023849880322813988, + 0.017550455406308174, + 0.0072030057199299335, + 0.00924011878669262, + 0.004704069346189499, + 0.014682251028716564, + 0.04217943921685219, + 0.0075517683289945126, + 0.023193305358290672, + 0.007164415903389454, + 0.005830009002238512, + 0.009614947251975536, + 0.007039643824100494, + 0.014738990925252438, + 0.005504005588591099, + 0.007482326123863459, + 0.00475333584472537, + 0.005357875023037195, + 0.0072214663960039616, + 0.006694640964269638, + 0.0075860763899981976, + 0.008300775662064552, + 0.005022893194109201, + 0.008912095800042152, + 0.005031663924455643, + 0.011547090485692024, + 0.01617850922048092, + 0.016063321381807327, + 0.01366408821195364, + 0.015981629490852356, + 0.011642453260719776, + 0.02116784267127514, + 0.012271923944354057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0789676308631897 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_tensor([1241, 1241])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_tensor([1241, 1241])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a869221adf14a1d88e86aace2d4465e08b913703 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_tensor([1241, 1241])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, top-down, extreme close-up maintains a fixed frame on a flat sheet of pristine aluminum foil resting on a dark, polished black countertop. A single, bright overhead light source creates a stark, high-contrast environment, casting brilliant, mirror-like highlights on the foil's smooth, cool-toned silver surface and faintly reflecting them in the countertop's mirror-like finish. Suddenly, an invisible, localized downward pressure is deliberately applied to the center of the foil, causing the smooth surface to collapse into a gentle depression. This initial deformation instantly triggers a violent chain reaction; the surrounding material deforms under the pressure, erupting into a complex, spiderweb-like network of sharp, permanent creases. The foil wrinkles and folds in on itself, then enters a new, permanently deformed state. The camera captures how the newly formed ridges and angular valleys of the metal now catch the intense light, creating a chaotic, shimmering pattern of sharp reflections and deep shadows, contrasting the memory of the original smooth surface with the final, wrinkled texture, creating an atmosphere of tension, abstraction, and visual drama.", + "seed": 13759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13759_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.25346848510671405, + "vlm_logprob": -0.25346848510671405, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.25346848510671405, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2374320030212402 + ], + "opposite_logprob_tokens": [ + -0.3624320328235626 + ], + "logprob_reward": -1.2374320030212402, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2374320030212402, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09405984729528427, + 0.06613922119140625, + 0.1712072342634201, + 0.2252434641122818, + 0.12208971381187439, + 0.11457755416631699, + 0.08863366395235062, + 0.043062783777713776, + 0.04273694008588791, + 0.03224962204694748 + ], + "credit_map": [ + 0.015471082180738449, + 0.006197670474648476, + 0.005268057808279991, + 0.007548904977738857, + 0.009849105030298233, + 0.023668993264436722, + 0.009682271629571915, + 0.016373757272958755, + 0.011132722720503807, + 0.005397598259150982, + 0.005110729951411486, + 0.006997261196374893, + 0.0060112858191132545, + 0.015092439018189907, + 0.005395007319748402, + 0.011002167128026485, + 0.030361875891685486, + 0.017069779336452484, + 0.024602316319942474, + 0.026005802676081657, + 0.024104662239551544, + 0.012459957972168922, + 0.013968286104500294, + 0.022634537890553474, + 0.04317893832921982, + 0.026044949889183044, + 0.032057393342256546, + 0.015688804909586906, + 0.02743544802069664, + 0.021398987621068954, + 0.03003793954849243, + 0.029400985687971115, + 0.0173984132707119, + 0.019119875505566597, + 0.017384378239512444, + 0.008901651948690414, + 0.028612064197659492, + 0.011845570988953114, + 0.012368123047053814, + 0.006459630560129881, + 0.013270063325762749, + 0.022571012377738953, + 0.011282228864729404, + 0.010550417937338352, + 0.014996061101555824, + 0.01258972566574812, + 0.015214859507977962, + 0.014103168621659279, + 0.008453208953142166, + 0.011459261178970337, + 0.015087454579770565, + 0.020419372245669365, + 0.0070326440036296844, + 0.005092372186481953, + 0.012369646690785885, + 0.008719704113900661, + 0.005471103359013796, + 0.006834626663476229, + 0.0063545978628098965, + 0.005617043003439903, + 0.003553001442924142, + 0.0028796768747270107, + 0.006424328312277794, + 0.00592840975150466, + 0.004079451318830252, + 0.002475917572155595, + 0.0036769607104361057, + 0.00368478917516768, + 0.0016264956211671233, + 0.007580141071230173, + 0.007970123551785946, + 0.011643054895102978, + 0.003402812872081995, + 0.0036048211622983217, + 0.004509488586336374, + 0.003144702408462763, + 0.00452465983107686, + 0.005126107484102249, + 0.0028478389140218496, + 0.00508918846026063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -1.2374320030212402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020625749602913857 + ], + "opposite_logprob_tokens": [ + -3.8956258296966553 + ], + "logprob_reward": -0.020625749602913857, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020625749602913857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16785208880901337, + 0.07457999140024185, + 0.19562982022762299, + 0.07767300307750702, + 0.1038653776049614, + 0.11062859743833542, + 0.06960601359605789, + 0.031084943562746048, + 0.12430492043495178, + 0.04477526992559433 + ], + "credit_map": [ + 0.024273963645100594, + 0.01344043668359518, + 0.015108248218894005, + 0.030690019950270653, + 0.01334280427545309, + 0.03116634115576744, + 0.012289222329854965, + 0.02754104696214199, + 0.01113260816782713, + 0.0064645325765013695, + 0.005213957745581865, + 0.016833966597914696, + 0.005435708444565535, + 0.011358908377587795, + 0.006752754095941782, + 0.011387553066015244, + 0.034875769168138504, + 0.018597658723592758, + 0.03505905345082283, + 0.018419910222291946, + 0.022316329181194305, + 0.016817409545183182, + 0.020831089466810226, + 0.028712602332234383, + 0.013185277581214905, + 0.010058553889393806, + 0.014733625575900078, + 0.005638748873025179, + 0.014206334948539734, + 0.007828920148313046, + 0.004899602383375168, + 0.007121945731341839, + 0.014164724387228489, + 0.019119733944535255, + 0.008448957465589046, + 0.007341123651713133, + 0.02546285279095173, + 0.012462740764021873, + 0.010411643423140049, + 0.006453602109104395, + 0.02226385660469532, + 0.02478441223502159, + 0.00941457413136959, + 0.009904099628329277, + 0.013189094141125679, + 0.009037641808390617, + 0.014103570953011513, + 0.007931359112262726, + 0.00664185918867588, + 0.005721569061279297, + 0.008971468545496464, + 0.02088720165193081, + 0.0049790991470217705, + 0.0034998911432921886, + 0.012639283202588558, + 0.006265643052756786, + 0.004547403659671545, + 0.005290536675602198, + 0.0029705544002354145, + 0.0044908239506185055, + 0.002862377557903528, + 0.0023314030840992928, + 0.0047389124520123005, + 0.0038529294542968273, + 0.01906556263566017, + 0.021738750860095024, + 0.0055723111145198345, + 0.006376217119395733, + 0.004805892240256071, + 0.04768812656402588, + 0.0043796650134027, + 0.014678399078547955, + 0.0044897738844156265, + 0.005863427184522152, + 0.00757801067084074, + 0.004832619335502386, + 0.005499868653714657, + 0.0055892979726195335, + 0.003769113915041089, + 0.0071531604044139385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.020625749602913857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017263285117223859 + ], + "opposite_logprob_tokens": [ + -6.376726150512695 + ], + "logprob_reward": -0.0017263285117223859, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017263285117223859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10543413460254669, + 0.05869295075535774, + 0.1642359346151352, + 0.1710219383239746, + 0.1375115066766739, + 0.097104012966156, + 0.09507250785827637, + 0.035027772188186646, + 0.0900241956114769, + 0.04587502405047417 + ], + "credit_map": [ + 0.015795262530446053, + 0.008800885640084743, + 0.007998191751539707, + 0.022534018382430077, + 0.011830865405499935, + 0.014602424576878548, + 0.010950462892651558, + 0.012922021560370922, + 0.010313035920262337, + 0.00537743279710412, + 0.005112601909786463, + 0.0054254992865026, + 0.005422426853328943, + 0.012406952679157257, + 0.004302497021853924, + 0.010332505218684673, + 0.02768557146191597, + 0.017318928614258766, + 0.026479952037334442, + 0.0311172716319561, + 0.01688578724861145, + 0.01119963824748993, + 0.0112508749589324, + 0.022297905758023262, + 0.03247210010886192, + 0.02013379894196987, + 0.024995125830173492, + 0.011888371780514717, + 0.022242560982704163, + 0.016939962282776833, + 0.02092001587152481, + 0.021430013701319695, + 0.019152091816067696, + 0.02523035742342472, + 0.015176436863839626, + 0.009887352585792542, + 0.03316934034228325, + 0.01379060372710228, + 0.01395405363291502, + 0.007151269353926182, + 0.01685549132525921, + 0.02116404101252556, + 0.008843185380101204, + 0.009062113240361214, + 0.011683005839586258, + 0.010015825740993023, + 0.012073406018316746, + 0.007406953256577253, + 0.007903787307441235, + 0.012624693103134632, + 0.01554935798048973, + 0.0220355074852705, + 0.007904274389147758, + 0.005242145620286465, + 0.014883768744766712, + 0.00892898254096508, + 0.004002778325229883, + 0.005790914874523878, + 0.00448786374181509, + 0.004866252653300762, + 0.0028949270490556955, + 0.0022397073917090893, + 0.0059769004583358765, + 0.004768429324030876, + 0.013622838072478771, + 0.01051739975810051, + 0.004495266824960709, + 0.004774019122123718, + 0.003192518139258027, + 0.03244423121213913, + 0.0051802415400743484, + 0.015797680243849754, + 0.006494091358035803, + 0.005712323356419802, + 0.007246360182762146, + 0.005270680412650108, + 0.004938782658427954, + 0.00586150586605072, + 0.00402041943743825, + 0.0063308654353022575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.0017263285117223859 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744714453816414 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744714453816414, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744714453816414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10883039236068726, + 0.04497984051704407, + 0.05656314641237259, + 0.12928663194179535, + 0.10592766851186752, + 0.09782258421182632, + 0.1771610975265503, + 0.04434051737189293, + 0.1965511590242386, + 0.038536958396434784 + ], + "credit_map": [ + 0.018997222185134888, + 0.0073449742048978806, + 0.008100158534944057, + 0.013882642611861229, + 0.00838058814406395, + 0.02484319917857647, + 0.006736621726304293, + 0.02054498717188835, + 0.007751903962343931, + 0.003056664252653718, + 0.0028120398055762053, + 0.0028168882708996534, + 0.0035627936013042927, + 0.010487161576747894, + 0.0046605938114225864, + 0.009831798262894154, + 0.010015081614255905, + 0.005159720778465271, + 0.012688644230365753, + 0.005601358134299517, + 0.0061845798045396805, + 0.004881467204540968, + 0.004194125533103943, + 0.00783817283809185, + 0.025550158694386482, + 0.015198706649243832, + 0.019920438528060913, + 0.008891084231436253, + 0.01745360903441906, + 0.012756223790347576, + 0.015008405782282352, + 0.014508017338812351, + 0.014245020225644112, + 0.024222688749432564, + 0.013956491835415363, + 0.008934861049056053, + 0.016562046483159065, + 0.011147848330438137, + 0.010082278400659561, + 0.006776438560336828, + 0.010987489484250546, + 0.013682514429092407, + 0.008924389258027077, + 0.009621181525290012, + 0.013388125225901604, + 0.011073012836277485, + 0.01786501519382, + 0.012280861847102642, + 0.013676383532583714, + 0.017838623374700546, + 0.025967007502913475, + 0.052085280418395996, + 0.013021964579820633, + 0.009132775478065014, + 0.03149868920445442, + 0.013940388336777687, + 0.004767340142279863, + 0.009488813579082489, + 0.005860880482941866, + 0.0067308018915355206, + 0.003924552816897631, + 0.0029671085067093372, + 0.006465617101639509, + 0.004135408904403448, + 0.02084491029381752, + 0.0584367997944355, + 0.01104872114956379, + 0.011689533479511738, + 0.011817799881100655, + 0.06100424751639366, + 0.007466420065611601, + 0.014242738485336304, + 0.005590730346739292, + 0.0052850451320409775, + 0.005352166015654802, + 0.0038025539834052324, + 0.004171620588749647, + 0.005709708668291569, + 0.003143721492961049, + 0.005481414962559938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.006744714453816414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008136299438774586 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008136299438774586, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008136299438774586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051094766706228256, + 0.05798083171248436, + 0.13803011178970337, + 0.14403291046619415, + 0.18309840559959412, + 0.12642470002174377, + 0.10273870825767517, + 0.06538166105747223, + 0.08926673978567123, + 0.04195115715265274 + ], + "credit_map": [ + 0.008324332535266876, + 0.004801688715815544, + 0.003995668608695269, + 0.005358382128179073, + 0.007208249066025019, + 0.007066444959491491, + 0.0069953943602740765, + 0.007344608660787344, + 0.010634655132889748, + 0.004249007906764746, + 0.004603780806064606, + 0.003045140067115426, + 0.004845909308642149, + 0.014439062215387821, + 0.003937650006264448, + 0.012225629761815071, + 0.024392489343881607, + 0.01456542406231165, + 0.01838749647140503, + 0.02251219004392624, + 0.0204678513109684, + 0.010259521193802357, + 0.010698623955249786, + 0.01674652099609375, + 0.027347926050424576, + 0.01766469143331051, + 0.02143130823969841, + 0.01016991026699543, + 0.01871335133910179, + 0.014380394481122494, + 0.017432335764169693, + 0.016892988234758377, + 0.025952637195587158, + 0.0313301645219326, + 0.025471922010183334, + 0.013445470482110977, + 0.04137323796749115, + 0.017816884443163872, + 0.0180361345410347, + 0.009671946987509727, + 0.017299892380833626, + 0.02769159898161888, + 0.011058313772082329, + 0.01082740817219019, + 0.01509024016559124, + 0.012818423099815845, + 0.016254998743534088, + 0.015383823774755001, + 0.008775151334702969, + 0.01327925082296133, + 0.018619515001773834, + 0.023379571735858917, + 0.008735192008316517, + 0.006397278513759375, + 0.0136155029758811, + 0.00993724912405014, + 0.007974113337695599, + 0.012190544977784157, + 0.00963505543768406, + 0.008969736285507679, + 0.006517854984849691, + 0.005767538212239742, + 0.007350820116698742, + 0.006975997239351273, + 0.010355588048696518, + 0.016462864354252815, + 0.006765802390873432, + 0.006535768508911133, + 0.00564326299354434, + 0.024655262008309364, + 0.009820369072258472, + 0.009027828462421894, + 0.0058259982615709305, + 0.005198456346988678, + 0.005958165042102337, + 0.004207021091133356, + 0.0041856649331748486, + 0.006972862873226404, + 0.0035032567102462053, + 0.006099729798734188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.0008136299438774586 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down, extreme close-up maintains a fixed frame on a flat sheet of pristine aluminum foil resting on a dark, polished black countertop. A single, bright overhead light source creates a stark, high-contrast environment, casting brilliant, mirror-like highlights on the foil's smooth, cool-toned silver surface and faintly reflecting them in the countertop's mirror-like finish. Suddenly, an invisible, localized downward pressure is deliberately applied to the center of the foil, causing the smooth surface to collapse into a gentle depression. This initial deformation instantly triggers a violent chain reaction; the surrounding material deforms under the pressure, erupting into a complex, spiderweb-like network of sharp, permanent creases. The foil wrinkles and folds in on itself, then enters a new, permanently deformed state. The camera captures how the newly formed ridges and angular valleys of the metal now catch the intense light, creating a chaotic, shimmering pattern of sharp reflections and deep shadows, contrasting the memory of the original smooth surface with the final, wrinkled texture, creating an atmosphere of tension, abstraction, and visual drama.", + "seed": 13760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13760_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1395609048893675, + "vlm_logprob": -0.1395609048893675, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.1395609048893675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6404043436050415 + ], + "opposite_logprob_tokens": [ + -0.7654043436050415 + ], + "logprob_reward": -0.6404043436050415, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6404043436050415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13512475788593292, + 0.09251462668180466, + 0.11749452352523804, + 0.08636699616909027, + 0.20262210071086884, + 0.06061011552810669, + 0.056011833250522614, + 0.03883649408817291, + 0.096100814640522, + 0.11431772261857986 + ], + "credit_map": [ + 0.015513083897531033, + 0.01809140108525753, + 0.023947538807988167, + 0.011935209855437279, + 0.022365134209394455, + 0.013302627950906754, + 0.015245556831359863, + 0.01472420059144497, + 0.01559691783040762, + 0.009892544709146023, + 0.014921684749424458, + 0.011653699912130833, + 0.010735837742686272, + 0.008877396583557129, + 0.009616990573704243, + 0.011219543404877186, + 0.014382547698915005, + 0.012171206064522266, + 0.0166951734572649, + 0.018681740388274193, + 0.01863308623433113, + 0.010696404613554478, + 0.013413568958640099, + 0.012820788659155369, + 0.010230820626020432, + 0.006412632763385773, + 0.011740021407604218, + 0.007927018217742443, + 0.006231899373233318, + 0.014822798781096935, + 0.011538638733327389, + 0.017463164404034615, + 0.03679932281374931, + 0.013993979431688786, + 0.024038469418883324, + 0.01707945019006729, + 0.01827576942741871, + 0.024060864001512527, + 0.020993642508983612, + 0.047380585223436356, + 0.008242621086537838, + 0.0049120052717626095, + 0.008748541586101055, + 0.005120322108268738, + 0.008101287297904491, + 0.00594685785472393, + 0.00803384929895401, + 0.01150462869554758, + 0.00708317244425416, + 0.006087350193411112, + 0.0048415884375572205, + 0.004536863882094622, + 0.01554903294891119, + 0.0052533382549881935, + 0.006536722648888826, + 0.006123760715126991, + 0.004776435438543558, + 0.003941758535802364, + 0.004037010483443737, + 0.0037229773588478565, + 0.0038624871522188187, + 0.004488724283874035, + 0.009671645238995552, + 0.004335454665124416, + 0.01158563606441021, + 0.011138521134853363, + 0.012567877769470215, + 0.019086221233010292, + 0.007094854488968849, + 0.009494839236140251, + 0.01797626167535782, + 0.007156597916036844, + 0.013064061291515827, + 0.012086817063391209, + 0.016533417627215385, + 0.02787296287715435, + 0.012663403525948524, + 0.0050015440210700035, + 0.019095564261078835, + 0.007999952882528305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.6404043436050415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868509620428085 + ], + "opposite_logprob_tokens": [ + -3.048685073852539 + ], + "logprob_reward": -0.04868509620428085, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868509620428085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03772513568401337, + 0.05001465231180191, + 0.06221553310751915, + 0.06164262443780899, + 0.3526860773563385, + 0.06537851691246033, + 0.12747050821781158, + 0.0742277055978775, + 0.05371882766485214, + 0.11492043733596802 + ], + "credit_map": [ + 0.005055570509284735, + 0.004552238155156374, + 0.005068791564553976, + 0.0035009081475436687, + 0.004852126352488995, + 0.006024555768817663, + 0.0036530743818730116, + 0.005017869174480438, + 0.008066914044320583, + 0.005585371050983667, + 0.008175141178071499, + 0.0048497566021978855, + 0.007095922250300646, + 0.005678527522832155, + 0.005111550446599722, + 0.00545146781951189, + 0.006713345181196928, + 0.006922407075762749, + 0.00785236619412899, + 0.007345274090766907, + 0.010538196191191673, + 0.006513391155749559, + 0.006790816783905029, + 0.009539731778204441, + 0.006546894554048777, + 0.004782057832926512, + 0.008546531200408936, + 0.007075150031596422, + 0.004762307275086641, + 0.00969215203076601, + 0.008873719722032547, + 0.011363810859620571, + 0.06381414830684662, + 0.024403171613812447, + 0.04058001562952995, + 0.025532850995659828, + 0.04499874264001846, + 0.04175298288464546, + 0.03391788899898529, + 0.07768625766038895, + 0.0064382413402199745, + 0.005419870838522911, + 0.011911910027265549, + 0.009144854731857777, + 0.006268325727432966, + 0.006094064097851515, + 0.012394565157592297, + 0.007706681732088327, + 0.01683974079787731, + 0.012982162646949291, + 0.01153681892901659, + 0.00890867505222559, + 0.03458629176020622, + 0.01261583250015974, + 0.013537420891225338, + 0.01646355539560318, + 0.008668945170938969, + 0.010958422906696796, + 0.010868757963180542, + 0.008948923088610172, + 0.008987484499812126, + 0.006092065945267677, + 0.009823419153690338, + 0.00987968035042286, + 0.006233208812773228, + 0.005100918468087912, + 0.006495562382042408, + 0.013225562870502472, + 0.005948873236775398, + 0.0029101299587637186, + 0.008343671448528767, + 0.0054608965292572975, + 0.012186397798359394, + 0.012998856604099274, + 0.020892323926091194, + 0.024108530953526497, + 0.013589795678853989, + 0.005386312026530504, + 0.017871633172035217, + 0.007886584848165512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.04868509620428085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015195977175608277 + ], + "opposite_logprob_tokens": [ + -6.501519680023193 + ], + "logprob_reward": -0.0015195977175608277, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015195977175608277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10951583832502365, + 0.132078617811203, + 0.08593552559614182, + 0.14671851694583893, + 0.14643453061580658, + 0.03671348839998245, + 0.10930220782756805, + 0.08508974313735962, + 0.04610423371195793, + 0.10210732370615005 + ], + "credit_map": [ + 0.013634822331368923, + 0.012904386036098003, + 0.01723301224410534, + 0.011274101212620735, + 0.014909033663570881, + 0.012527606450021267, + 0.013984805904328823, + 0.013048077002167702, + 0.01432974636554718, + 0.01543995551764965, + 0.019977601245045662, + 0.014064842835068703, + 0.020113874226808548, + 0.013841994106769562, + 0.017651159316301346, + 0.0166594497859478, + 0.009600730612874031, + 0.009351125918328762, + 0.012771036475896835, + 0.012711703777313232, + 0.013453347608447075, + 0.008467566221952438, + 0.009509350173175335, + 0.01007066573947668, + 0.015868058428168297, + 0.01350738387554884, + 0.02425946108996868, + 0.020161034539341927, + 0.01302009541541338, + 0.016542406752705574, + 0.023215817287564278, + 0.020144259557127953, + 0.024025000631809235, + 0.010373900644481182, + 0.01560867577791214, + 0.013124347664415836, + 0.018687162548303604, + 0.017827974632382393, + 0.016470737755298615, + 0.030316734686493874, + 0.004707672167569399, + 0.004197436384856701, + 0.00523752998560667, + 0.0051925331354141235, + 0.0034631025046110153, + 0.003698464250192046, + 0.0049890480004251, + 0.005227703135460615, + 0.014330613426864147, + 0.012215817347168922, + 0.009065171703696251, + 0.006583169102668762, + 0.03304940462112427, + 0.010858172550797462, + 0.011089042760431767, + 0.012110820040106773, + 0.008250762708485126, + 0.014718909747898579, + 0.01237065251916647, + 0.009178008884191513, + 0.012295233085751534, + 0.006952039431780577, + 0.01054703164845705, + 0.010777096264064312, + 0.005002926103770733, + 0.0059666503220796585, + 0.007046398241072893, + 0.006791531108319759, + 0.004148802254348993, + 0.005682138726115227, + 0.007242944557219744, + 0.004222840536385775, + 0.014035668224096298, + 0.010095507837831974, + 0.0143965445458889, + 0.018874747678637505, + 0.011443103663623333, + 0.005499920342117548, + 0.018867358565330505, + 0.008894471451640129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.0015195977175608277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052567594684660435 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052567594684660435, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052567594684660435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08346939086914062, + 0.11608633399009705, + 0.15344831347465515, + 0.17563244700431824, + 0.08657745271921158, + 0.06328116357326508, + 0.08167001605033875, + 0.07347129285335541, + 0.04905790090560913, + 0.11730571836233139 + ], + "credit_map": [ + 0.009185190312564373, + 0.011083419434726238, + 0.012031001970171928, + 0.00822186004370451, + 0.013295733369886875, + 0.009355057962238789, + 0.009250801056623459, + 0.011046333238482475, + 0.01800435036420822, + 0.014906547032296658, + 0.017760198563337326, + 0.009980052709579468, + 0.01881551370024681, + 0.014181816950440407, + 0.009728474542498589, + 0.012709373608231544, + 0.01892123557627201, + 0.015470004640519619, + 0.020962590351700783, + 0.02290746010839939, + 0.02630004845559597, + 0.014188853092491627, + 0.01683039218187332, + 0.01786772347986698, + 0.01861807145178318, + 0.015985552221536636, + 0.02568878047168255, + 0.02692500501871109, + 0.014544305391609669, + 0.02216438390314579, + 0.028254825621843338, + 0.023451512679457664, + 0.012827875092625618, + 0.005952190607786179, + 0.011182505637407303, + 0.010789401829242706, + 0.00896480493247509, + 0.009020796976983547, + 0.011250804178416729, + 0.016589073464274406, + 0.007066160906106234, + 0.005464817397296429, + 0.007150136400014162, + 0.006979512982070446, + 0.007900956086814404, + 0.0076773944310843945, + 0.009096034802496433, + 0.011946151033043861, + 0.011860768310725689, + 0.008096574805676937, + 0.007495098747313023, + 0.005077071022242308, + 0.02145233564078808, + 0.008269043639302254, + 0.008676099590957165, + 0.01074302103370428, + 0.00698422035202384, + 0.008938995189964771, + 0.00919374730437994, + 0.008648182265460491, + 0.008127330802381039, + 0.00886326003819704, + 0.0133499251678586, + 0.009365629404783249, + 0.0055924332700669765, + 0.006924700923264027, + 0.00741004291921854, + 0.007446931675076485, + 0.0038504754193127155, + 0.005846092011779547, + 0.007496958132833242, + 0.004490270279347897, + 0.012025848031044006, + 0.01390163879841566, + 0.01594061590731144, + 0.028622474521398544, + 0.012457771226763725, + 0.005521694663912058, + 0.02052159421145916, + 0.00831407867372036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.0052567594684660435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019387274514883757 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019387274514883757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019387274514883757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06474345922470093, + 0.052478011697530746, + 0.04054555296897888, + 0.057905953377485275, + 0.12951430678367615, + 0.31976550817489624, + 0.14059029519557953, + 0.06439529359340668, + 0.055379468947649, + 0.07468219101428986 + ], + "credit_map": [ + 0.007515490520745516, + 0.009427301585674286, + 0.011368747800588608, + 0.006279470399022102, + 0.009145914576947689, + 0.0069340309128165245, + 0.006485205143690109, + 0.007587290834635496, + 0.008368765003979206, + 0.005944336298853159, + 0.008101069368422031, + 0.00573553703725338, + 0.006483438424766064, + 0.006104293745011091, + 0.0048952787183225155, + 0.006845288909971714, + 0.004132495727390051, + 0.0040864297188818455, + 0.004685836844146252, + 0.006680883467197418, + 0.005792565643787384, + 0.003789314767345786, + 0.005167736206203699, + 0.006210291292518377, + 0.006679126992821693, + 0.00484606996178627, + 0.008971345610916615, + 0.008646293543279171, + 0.004857046529650688, + 0.0065659936517477036, + 0.009290000423789024, + 0.008050074800848961, + 0.01627483032643795, + 0.008589426055550575, + 0.013722866773605347, + 0.014213203452527523, + 0.015807805582880974, + 0.015399176627397537, + 0.014677578583359718, + 0.0308294128626585, + 0.03098197840154171, + 0.02727247215807438, + 0.0611562617123127, + 0.03718285635113716, + 0.03331137076020241, + 0.02736104466021061, + 0.06442682445049286, + 0.03807268664240837, + 0.01566966064274311, + 0.018476467579603195, + 0.013756223022937775, + 0.009795770049095154, + 0.039585623890161514, + 0.011875339783728123, + 0.014371122233569622, + 0.017060091719031334, + 0.005947928410023451, + 0.007101719733327627, + 0.007660785224288702, + 0.008259477093815804, + 0.00877266377210617, + 0.0070670838467776775, + 0.011951119638979435, + 0.007634513545781374, + 0.007268261164426804, + 0.007762878201901913, + 0.006910665892064571, + 0.010169719345867634, + 0.0040605198591947556, + 0.006448909640312195, + 0.008535264059901237, + 0.004223247524350882, + 0.010460270568728447, + 0.00760270282626152, + 0.009443813934922218, + 0.011100930161774158, + 0.009681656956672668, + 0.004206681624054909, + 0.015584156848490238, + 0.00660196878015995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0019387274514883757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down, extreme close-up maintains a fixed frame on a flat sheet of pristine aluminum foil resting on a dark, polished black countertop. A single, bright overhead light source creates a stark, high-contrast environment, casting brilliant, mirror-like highlights on the foil's smooth, cool-toned silver surface and faintly reflecting them in the countertop's mirror-like finish. Suddenly, an invisible, localized downward pressure is deliberately applied to the center of the foil, causing the smooth surface to collapse into a gentle depression. This initial deformation instantly triggers a violent chain reaction; the surrounding material deforms under the pressure, erupting into a complex, spiderweb-like network of sharp, permanent creases. The foil wrinkles and folds in on itself, then enters a new, permanently deformed state. The camera captures how the newly formed ridges and angular valleys of the metal now catch the intense light, creating a chaotic, shimmering pattern of sharp reflections and deep shadows, contrasting the memory of the original smooth surface with the final, wrinkled texture, creating an atmosphere of tension, abstraction, and visual drama.", + "seed": 13761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13761_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12731661666184663, + "vlm_logprob": -0.12731661666184663, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.12731661666184663, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.590623140335083 + ], + "opposite_logprob_tokens": [ + -0.840623140335083 + ], + "logprob_reward": -0.590623140335083, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.590623140335083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29011601209640503, + 0.10298766195774078, + 0.044210728257894516, + 0.21234792470932007, + 0.09683050215244293, + 0.079530268907547, + 0.03470868989825249, + 0.0487992987036705, + 0.025449290871620178, + 0.0650196298956871 + ], + "credit_map": [ + 0.036152154207229614, + 0.03231062367558479, + 0.030660495162010193, + 0.02711978554725647, + 0.03195761889219284, + 0.0581221841275692, + 0.03554927557706833, + 0.03824387863278389, + 0.011090721003711224, + 0.007852289825677872, + 0.01405412144958973, + 0.011456803418695927, + 0.007580376695841551, + 0.018563609570264816, + 0.01869492419064045, + 0.013694819062948227, + 0.004641515668481588, + 0.00564773753285408, + 0.004882985260337591, + 0.004924322012811899, + 0.004155609291046858, + 0.008058893494307995, + 0.005596308037638664, + 0.006303360685706139, + 0.023385198786854744, + 0.025906827300786972, + 0.031533133238554, + 0.02720049023628235, + 0.019114835187792778, + 0.03922154754400253, + 0.01803627982735634, + 0.0279496181756258, + 0.009395643137395382, + 0.008432211354374886, + 0.014461026526987553, + 0.009914057329297066, + 0.011383265256881714, + 0.019863728433847427, + 0.011318584904074669, + 0.012061987072229385, + 0.013096781447529793, + 0.006835067179054022, + 0.011246668174862862, + 0.0070952423848211765, + 0.006335272453725338, + 0.014794734306633472, + 0.006984998472034931, + 0.013141513802111149, + 0.004547525197267532, + 0.003751941490918398, + 0.004560327157378197, + 0.005250847432762384, + 0.0035283053293824196, + 0.004583018831908703, + 0.004692481830716133, + 0.003794246120378375, + 0.00707416795194149, + 0.004169349558651447, + 0.006627925205975771, + 0.006027356721460819, + 0.005159006454050541, + 0.008335213176906109, + 0.0047304509207606316, + 0.006675831042230129, + 0.0035743704065680504, + 0.0025382607709616423, + 0.003247700398787856, + 0.003126687603071332, + 0.004057673737406731, + 0.0027904221788048744, + 0.002724091988056898, + 0.0033900854177773, + 0.0052038743160665035, + 0.01268863771110773, + 0.008647384122014046, + 0.006692875176668167, + 0.016184045001864433, + 0.004988409578800201, + 0.005325695499777794, + 0.005288709886372089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.590623140335083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813987225294113 + ], + "opposite_logprob_tokens": [ + -3.288139820098877 + ], + "logprob_reward": -0.03813987225294113, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813987225294113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3152366280555725, + 0.24248163402080536, + 0.06650559604167938, + 0.07268182188272476, + 0.0912739709019661, + 0.05138145759701729, + 0.02950058877468109, + 0.031587909907102585, + 0.04744790494441986, + 0.05190247669816017 + ], + "credit_map": [ + 0.036874134093523026, + 0.02746724523603916, + 0.038151517510414124, + 0.03289244696497917, + 0.032327111810445786, + 0.06655964255332947, + 0.041995491832494736, + 0.03896906226873398, + 0.028667166829109192, + 0.020017370581626892, + 0.03207264468073845, + 0.02825726382434368, + 0.0222482830286026, + 0.0453249029815197, + 0.034461796283721924, + 0.031432218849658966, + 0.008479689247906208, + 0.006964345928281546, + 0.009031003341078758, + 0.006179979536682367, + 0.007009963970631361, + 0.011689192615449429, + 0.008408976718783379, + 0.00874244887381792, + 0.007931575179100037, + 0.01093184296041727, + 0.008468394167721272, + 0.010426334105432034, + 0.006372059695422649, + 0.011175737716257572, + 0.0066953119821846485, + 0.01068057119846344, + 0.011286022141575813, + 0.009264587424695492, + 0.013183566741645336, + 0.007073850836604834, + 0.010635645128786564, + 0.01995895430445671, + 0.007077760063111782, + 0.012793589383363724, + 0.008209391497075558, + 0.004496719688177109, + 0.005897119175642729, + 0.0036910560447722673, + 0.0048973048105835915, + 0.013563026674091816, + 0.004272818099707365, + 0.0063540213741362095, + 0.004687730688601732, + 0.0027979861479252577, + 0.0028463113121688366, + 0.004748742561787367, + 0.0021011095959693193, + 0.004352943040430546, + 0.004131754860281944, + 0.0038340131286531687, + 0.0037941173650324345, + 0.00276491092517972, + 0.003882634686306119, + 0.003930282779037952, + 0.0039041435811668634, + 0.006124903913587332, + 0.002885066205635667, + 0.004301851615309715, + 0.006745494436472654, + 0.005341308191418648, + 0.006677623838186264, + 0.006204855628311634, + 0.006115158088505268, + 0.005036983639001846, + 0.005288592539727688, + 0.006037891376763582, + 0.004365445114672184, + 0.007513382472097874, + 0.006545526906847954, + 0.005662655923515558, + 0.012591416016221046, + 0.003221612423658371, + 0.005317214410752058, + 0.0066852266900241375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03813987225294113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010519453790038824 + ], + "opposite_logprob_tokens": [ + -6.876051902770996 + ], + "logprob_reward": -0.0010519453790038824, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010519453790038824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29297587275505066, + 0.17312441766262054, + 0.0773870125412941, + 0.17281027138233185, + 0.07959481328725815, + 0.03822989761829376, + 0.03483601287007332, + 0.04817183315753937, + 0.03598400577902794, + 0.046885859221220016 + ], + "credit_map": [ + 0.03311173990368843, + 0.027681833133101463, + 0.03919751197099686, + 0.030240779742598534, + 0.026562148705124855, + 0.05653386190533638, + 0.030775638297200203, + 0.048872336745262146, + 0.017819197848439217, + 0.013708329759538174, + 0.024266231805086136, + 0.01782272197306156, + 0.013467435725033283, + 0.03372812271118164, + 0.028810473158955574, + 0.023501913994550705, + 0.007416114676743746, + 0.008122616447508335, + 0.00949767604470253, + 0.008189829997718334, + 0.007277415134012699, + 0.015087991952896118, + 0.010773321613669395, + 0.011022045277059078, + 0.016293182969093323, + 0.025573791936039925, + 0.0191880464553833, + 0.02287743054330349, + 0.020261643454432487, + 0.029171103611588478, + 0.01840708963572979, + 0.021037979051470757, + 0.008389386348426342, + 0.007652931846678257, + 0.012635717168450356, + 0.009849702008068562, + 0.0077524431981146336, + 0.011881493963301182, + 0.011117459274828434, + 0.010315682739019394, + 0.006379505153745413, + 0.0035189217887818813, + 0.004749676212668419, + 0.003664933377876878, + 0.0035413203295320272, + 0.006878559943288565, + 0.0037720119580626488, + 0.005724968388676643, + 0.003620854113250971, + 0.0030090631917119026, + 0.004113770555704832, + 0.00811594258993864, + 0.0031324659939855337, + 0.0037663797847926617, + 0.0052526844665408134, + 0.0038248514756560326, + 0.006775151938199997, + 0.004268321208655834, + 0.005954509135335684, + 0.007567279040813446, + 0.005283909384161234, + 0.00737500935792923, + 0.005888004787266254, + 0.005059645511209965, + 0.006451754365116358, + 0.0027646489907056093, + 0.005662732291966677, + 0.0038305690977722406, + 0.004610363859683275, + 0.0038433102890849113, + 0.003545010695233941, + 0.00527561642229557, + 0.004582221619784832, + 0.004399928264319897, + 0.013114193454384804, + 0.004398040939122438, + 0.009477158077061176, + 0.002641881350427866, + 0.004120283294469118, + 0.004152153618633747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0010519453790038824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955924745649099 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955924745649099, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955924745649099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.279258131980896, + 0.10047020018100739, + 0.06267895549535751, + 0.0987834557890892, + 0.13555383682250977, + 0.06479475647211075, + 0.05357584357261658, + 0.06997799128293991, + 0.06498754024505615, + 0.06991936266422272 + ], + "credit_map": [ + 0.040577758103609085, + 0.028483210131525993, + 0.033586129546165466, + 0.025433847680687904, + 0.03246210515499115, + 0.053367651998996735, + 0.030001439154148102, + 0.03534594550728798, + 0.01082634087651968, + 0.008679729886353016, + 0.013030117377638817, + 0.00926972646266222, + 0.008292097598314285, + 0.021449826657772064, + 0.015864085406064987, + 0.013058267533779144, + 0.0072447555139660835, + 0.006095132324844599, + 0.006247539073228836, + 0.00707479752600193, + 0.0054412721656262875, + 0.014572430402040482, + 0.00777787109836936, + 0.008225146681070328, + 0.010810931213200092, + 0.010707772336900234, + 0.013594739139080048, + 0.013911199755966663, + 0.008961527608335018, + 0.017308056354522705, + 0.010410451330244541, + 0.013078763149678707, + 0.015489127486944199, + 0.011115119792521, + 0.02442212775349617, + 0.010880029760301113, + 0.012099710293114185, + 0.031152799725532532, + 0.009850561618804932, + 0.020544344559311867, + 0.010378112085163593, + 0.005270648282021284, + 0.007831091061234474, + 0.005622985307127237, + 0.006028138101100922, + 0.013822413049638271, + 0.006382790859788656, + 0.00945857260376215, + 0.007134753279387951, + 0.006316159386187792, + 0.006353330798447132, + 0.009791430085897446, + 0.004871121607720852, + 0.008089612238109112, + 0.0050658052787184715, + 0.005953626241534948, + 0.010410563088953495, + 0.005791761912405491, + 0.007697400636970997, + 0.010114982724189758, + 0.008491210639476776, + 0.012489025481045246, + 0.008250116370618343, + 0.0067329187877476215, + 0.0076580229215323925, + 0.005543246399611235, + 0.007833155803382397, + 0.01034541055560112, + 0.009574421681463718, + 0.006601081229746342, + 0.009305281564593315, + 0.008126918226480484, + 0.008733629249036312, + 0.01072058454155922, + 0.008068915456533432, + 0.005622761324048042, + 0.014538798481225967, + 0.005684027448296547, + 0.0043396493420004845, + 0.012210991233587265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005955924745649099 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008122005965560675 + ], + "opposite_logprob_tokens": [ + -7.12581205368042 + ], + "logprob_reward": -0.0008122005965560675, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008122005965560675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2946460247039795, + 0.13494791090488434, + 0.06049171835184097, + 0.14651478826999664, + 0.10236971080303192, + 0.053356416523456573, + 0.04872070997953415, + 0.037231892347335815, + 0.04204335808753967, + 0.07967748492956161 + ], + "credit_map": [ + 0.04500274360179901, + 0.026867739856243134, + 0.03586621209979057, + 0.026280643418431282, + 0.02947632223367691, + 0.05992259457707405, + 0.031570423394441605, + 0.03965936601161957, + 0.014146512374281883, + 0.011997041292488575, + 0.018793325871229172, + 0.013116789981722832, + 0.010967981070280075, + 0.026518750935792923, + 0.021347079426050186, + 0.018060438334941864, + 0.006662828382104635, + 0.0077364156022667885, + 0.00611730944365263, + 0.005512875970453024, + 0.006318229716271162, + 0.012700214050710201, + 0.006907572038471699, + 0.008536272682249546, + 0.015460791066288948, + 0.01863512583076954, + 0.020250840112566948, + 0.020113615319132805, + 0.012421787716448307, + 0.02696346677839756, + 0.013122614473104477, + 0.019546542316675186, + 0.013137164525687695, + 0.009370946325361729, + 0.013402908109128475, + 0.011424383148550987, + 0.011057747527956963, + 0.021390452980995178, + 0.009466862305998802, + 0.01311925146728754, + 0.00811296422034502, + 0.005879778414964676, + 0.007487501483410597, + 0.005663479678332806, + 0.005047228652983904, + 0.00872291810810566, + 0.005485054105520248, + 0.006957489065825939, + 0.006243332289159298, + 0.004793843254446983, + 0.00584109453484416, + 0.007328306790441275, + 0.004268564283847809, + 0.006334497593343258, + 0.009129800833761692, + 0.004781271796673536, + 0.005057999398559332, + 0.003258287440985441, + 0.006063384003937244, + 0.004704809747636318, + 0.0043045589700341225, + 0.004811898805201054, + 0.003445786191150546, + 0.005585166625678539, + 0.008062135428190231, + 0.0028976404573768377, + 0.0072364360094070435, + 0.004234010353684425, + 0.005109149031341076, + 0.004015148617327213, + 0.004145210608839989, + 0.0063436273485422134, + 0.006268783938139677, + 0.014611952938139439, + 0.012792724184691906, + 0.007195155136287212, + 0.020210206508636475, + 0.006060909479856491, + 0.0058806356973946095, + 0.006657113321125507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0008122005965560675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down, extreme close-up maintains a fixed frame on a flat sheet of pristine aluminum foil resting on a dark, polished black countertop. A single, bright overhead light source creates a stark, high-contrast environment, casting brilliant, mirror-like highlights on the foil's smooth, cool-toned silver surface and faintly reflecting them in the countertop's mirror-like finish. Suddenly, an invisible, localized downward pressure is deliberately applied to the center of the foil, causing the smooth surface to collapse into a gentle depression. This initial deformation instantly triggers a violent chain reaction; the surrounding material deforms under the pressure, erupting into a complex, spiderweb-like network of sharp, permanent creases. The foil wrinkles and folds in on itself, then enters a new, permanently deformed state. The camera captures how the newly formed ridges and angular valleys of the metal now catch the intense light, creating a chaotic, shimmering pattern of sharp reflections and deep shadows, contrasting the memory of the original smooth surface with the final, wrinkled texture, creating an atmosphere of tension, abstraction, and visual drama.", + "seed": 13762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_7/index_1241_seed_13762_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.216752433241345, + "vlm_logprob": -0.216752433241345, + "vlm_yes_no_margin": 4.575000011920929, + "vlm_token_credit": -0.216752433241345, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0650320053100586 + ], + "opposite_logprob_tokens": [ + -0.44003206491470337 + ], + "logprob_reward": -1.0650320053100586, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0650320053100586, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06274443864822388, + 0.02607324719429016, + 0.03692265972495079, + 0.08713413029909134, + 0.13810865581035614, + 0.08710987865924835, + 0.08135498315095901, + 0.17246602475643158, + 0.085075244307518, + 0.2230108231306076 + ], + "credit_map": [ + 0.016047224402427673, + 0.0035478698555380106, + 0.00800636038184166, + 0.004607521928846836, + 0.004102834966033697, + 0.005329133477061987, + 0.003483050037175417, + 0.01762043684720993, + 0.005975689738988876, + 0.002244119066745043, + 0.00372274499386549, + 0.0016092044534161687, + 0.002046978799626231, + 0.002602881286293268, + 0.001425713300704956, + 0.006445913575589657, + 0.00594673864543438, + 0.0026183503214269876, + 0.0060274130664765835, + 0.0036268001422286034, + 0.00652614189311862, + 0.003128072014078498, + 0.0033483593724668026, + 0.0057007805444300175, + 0.011980398558080196, + 0.011076376773416996, + 0.016949009150266647, + 0.008112465031445026, + 0.012474285438656807, + 0.007825841195881367, + 0.0066972291097044945, + 0.012018520385026932, + 0.03273394703865051, + 0.011564988642930984, + 0.01194986142218113, + 0.03719506412744522, + 0.008558393456041813, + 0.013637324795126915, + 0.012783217243850231, + 0.009685845114290714, + 0.016151199117302895, + 0.00920018833130598, + 0.01191290095448494, + 0.01146648544818163, + 0.009600001387298107, + 0.012454917654395103, + 0.006284772418439388, + 0.010039403103291988, + 0.009951699525117874, + 0.011300845071673393, + 0.010944871231913567, + 0.008136922493577003, + 0.01414310559630394, + 0.009575916454195976, + 0.007695851381868124, + 0.009605761617422104, + 0.019876403734087944, + 0.020945675671100616, + 0.02700553834438324, + 0.022317256778478622, + 0.02521427348256111, + 0.0173567533493042, + 0.01808810792863369, + 0.021661996841430664, + 0.008098973892629147, + 0.013088196516036987, + 0.010870356112718582, + 0.011440797708928585, + 0.009895972907543182, + 0.007629855535924435, + 0.009357152506709099, + 0.014693924225866795, + 0.024704590439796448, + 0.03085014410316944, + 0.019792590290308, + 0.01856626756489277, + 0.03390035405755043, + 0.02141096070408821, + 0.03198155760765076, + 0.04180433973670006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -1.0650320053100586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011109166778624058 + ], + "opposite_logprob_tokens": [ + -4.511109352111816 + ], + "logprob_reward": -0.011109166778624058, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011109166778624058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12819141149520874, + 0.03616898134350777, + 0.04114384204149246, + 0.06814518570899963, + 0.25103333592414856, + 0.09632327407598495, + 0.09219645708799362, + 0.06320339441299438, + 0.10515149682760239, + 0.11844265460968018 + ], + "credit_map": [ + 0.039542585611343384, + 0.006003217771649361, + 0.014611697755753994, + 0.00789718609303236, + 0.006321545224636793, + 0.01029529981315136, + 0.005183631554245949, + 0.038336243480443954, + 0.009182294830679893, + 0.002320254221558571, + 0.00504207843914628, + 0.0021908872295171022, + 0.004070771858096123, + 0.0030589366797357798, + 0.0018872208893299103, + 0.008416538126766682, + 0.0055712806060910225, + 0.0031301085837185383, + 0.0059241438284516335, + 0.0037991225253790617, + 0.007932689040899277, + 0.0033958610147237778, + 0.004124980419874191, + 0.007265658583492041, + 0.008609794080257416, + 0.009315341711044312, + 0.011010373942553997, + 0.006286806892603636, + 0.012676112353801727, + 0.006550133228302002, + 0.005304654594510794, + 0.008391965180635452, + 0.06178034469485283, + 0.02109246328473091, + 0.0225636288523674, + 0.06869641691446304, + 0.01443792600184679, + 0.024754460901021957, + 0.022668199613690376, + 0.015039886347949505, + 0.018257131800055504, + 0.010082908906042576, + 0.0132544981315732, + 0.013700979761779308, + 0.00994875654578209, + 0.014191364869475365, + 0.00570282619446516, + 0.011184810660779476, + 0.009300538338720798, + 0.013677980750799179, + 0.012095835991203785, + 0.009877911768853664, + 0.016751142218708992, + 0.008878972381353378, + 0.010011534206569195, + 0.011602545157074928, + 0.007035376038402319, + 0.007676281966269016, + 0.009979363530874252, + 0.008690296672284603, + 0.008655990473926067, + 0.006373070180416107, + 0.006904194131493568, + 0.007888821884989738, + 0.01001095399260521, + 0.0158968735486269, + 0.013807883486151695, + 0.01471799612045288, + 0.012730889022350311, + 0.00901335570961237, + 0.01187704224139452, + 0.017096497118473053, + 0.012723894789814949, + 0.016445213928818703, + 0.010081084445118904, + 0.009902245365083218, + 0.017296360805630684, + 0.011465746909379959, + 0.018135501071810722, + 0.02239261381328106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.011109166778624058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001725257490761578 + ], + "opposite_logprob_tokens": [ + -6.376725196838379 + ], + "logprob_reward": -0.001725257490761578, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001725257490761578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10382822901010513, + 0.02273424156010151, + 0.045369263738393784, + 0.08487050235271454, + 0.2908776104450226, + 0.10787202417850494, + 0.052314240485429764, + 0.09651074558496475, + 0.07325030863285065, + 0.12237288802862167 + ], + "credit_map": [ + 0.032283417880535126, + 0.004694574978202581, + 0.012178308330476284, + 0.005526453256607056, + 0.005588049069046974, + 0.008233883418142796, + 0.005156268365681171, + 0.03016727603971958, + 0.006292798090726137, + 0.0012954056728631258, + 0.0027340759988874197, + 0.0012466561747714877, + 0.0015064756153151393, + 0.0023639206774532795, + 0.0012724767439067364, + 0.006022431422024965, + 0.009995252825319767, + 0.0026990382466465235, + 0.0048147267661988735, + 0.0036747115664184093, + 0.0058864401653409, + 0.0036256068851798773, + 0.002738416427746415, + 0.011935069225728512, + 0.011530939489603043, + 0.011655406095087528, + 0.015273664146661758, + 0.009010551497340202, + 0.01172314491122961, + 0.007378321141004562, + 0.0071059707552194595, + 0.011192504316568375, + 0.07192359119653702, + 0.022937873378396034, + 0.02632296271622181, + 0.08150048553943634, + 0.016829274594783783, + 0.02853223867714405, + 0.026659434661269188, + 0.01617175154387951, + 0.018381204456090927, + 0.01109139621257782, + 0.015337658114731312, + 0.015851512551307678, + 0.01126869022846222, + 0.016311144456267357, + 0.007059688214212656, + 0.012570721097290516, + 0.0066052512265741825, + 0.00846602488309145, + 0.006321717519313097, + 0.005325291771441698, + 0.009503434412181377, + 0.006378460209816694, + 0.004446021746844053, + 0.005268044304102659, + 0.01180734857916832, + 0.01129996683448553, + 0.015126485377550125, + 0.012316799722611904, + 0.013355535455048084, + 0.009904125705361366, + 0.009721037931740284, + 0.012979448772966862, + 0.007150409277528524, + 0.012764308601617813, + 0.007459460292011499, + 0.008854304440319538, + 0.01005350612103939, + 0.006532776635140181, + 0.00767506891861558, + 0.012760471552610397, + 0.013358423486351967, + 0.016894355416297913, + 0.01021899189800024, + 0.010091864503920078, + 0.018436923623085022, + 0.0119398208335042, + 0.01865231618285179, + 0.022780202329158783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.001725257490761578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260791163891554 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260791163891554, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260791163891554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07654933631420135, + 0.028643611818552017, + 0.07333995401859283, + 0.07952895760536194, + 0.33916616439819336, + 0.09056250751018524, + 0.06551983952522278, + 0.059595249593257904, + 0.06963390856981277, + 0.11746055632829666 + ], + "credit_map": [ + 0.023094411939382553, + 0.0040511139668524265, + 0.008043192327022552, + 0.0050200498662889, + 0.004781673196703196, + 0.006962021812796593, + 0.0032302786130458117, + 0.02136659063398838, + 0.008980891667306423, + 0.0013443949865177274, + 0.003273017005994916, + 0.0013843871420249343, + 0.00175770022906363, + 0.002510841703042388, + 0.001182003179565072, + 0.008210374042391777, + 0.018262954428792, + 0.003911885898560286, + 0.007246157620102167, + 0.00443315040320158, + 0.008331920951604843, + 0.00599113991484046, + 0.004314352758228779, + 0.020848385989665985, + 0.011344770900905132, + 0.009908568114042282, + 0.015527076087892056, + 0.008143202401697636, + 0.010388772934675217, + 0.007126238662749529, + 0.006421868689358234, + 0.010668453760445118, + 0.08043384552001953, + 0.025922734290361404, + 0.03166228160262108, + 0.09361208975315094, + 0.0207795612514019, + 0.03477011248469353, + 0.031121809035539627, + 0.02086368389427662, + 0.01660437323153019, + 0.008593302220106125, + 0.012562630698084831, + 0.011173183098435402, + 0.010146024636924267, + 0.014793438836932182, + 0.0065320925787091255, + 0.010157465003430843, + 0.0074017285369336605, + 0.009872505441308022, + 0.0085093192756176, + 0.007401292212307453, + 0.01024753786623478, + 0.007449943106621504, + 0.006804655771702528, + 0.007832856848835945, + 0.006915637757629156, + 0.007071650121361017, + 0.009153037331998348, + 0.008506948128342628, + 0.007948912680149078, + 0.00618370994925499, + 0.006126808002591133, + 0.007688545156270266, + 0.007193515542894602, + 0.010493895970284939, + 0.007874454371631145, + 0.009048602543771267, + 0.00833850633352995, + 0.006149606313556433, + 0.007772780489176512, + 0.012762535363435745, + 0.013436008244752884, + 0.016374386847019196, + 0.009973716922104359, + 0.009638287127017975, + 0.018134061247110367, + 0.011201098561286926, + 0.017021186649799347, + 0.021681804209947586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.005260791163891554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down view maintaining a fixed frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a flat sheet of pristine aluminum foil resting on a dark, polished black countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bright overhead light source create high-contrast lighting with brilliant mirror-like highlights on the foil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foil undergo a deformation where a localized downward pressure causes it to collapse into a depression and then erupt into a complex network of sharp, permanent creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state of the foil show a wrinkled texture with chaotic, shimmering patterns of sharp reflections and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006349454633891582 + ], + "opposite_logprob_tokens": [ + -7.375635147094727 + ], + "logprob_reward": -0.0006349454633891582, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006349454633891582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06288360059261322, + 0.021687084808945656, + 0.0937759131193161, + 0.10104724764823914, + 0.211689755320549, + 0.0635918602347374, + 0.07199303060770035, + 0.14279639720916748, + 0.09977173805236816, + 0.13076339662075043 + ], + "credit_map": [ + 0.01847720891237259, + 0.0030521571170538664, + 0.00863424688577652, + 0.004145484883338213, + 0.0026729947421699762, + 0.005298776552081108, + 0.0027609255630522966, + 0.017841799184679985, + 0.005343163385987282, + 0.0012768531451001763, + 0.003055392298847437, + 0.0013754168758168817, + 0.0027620829641819, + 0.001863855984993279, + 0.0011825739638879895, + 0.004827745258808136, + 0.028986746445298195, + 0.0037340386770665646, + 0.00918267946690321, + 0.0034898342564702034, + 0.00612888066098094, + 0.007694324478507042, + 0.0039969845674932, + 0.03056243248283863, + 0.014807960018515587, + 0.012213638052344322, + 0.019441045820713043, + 0.009330451488494873, + 0.01323768962174654, + 0.009379372000694275, + 0.007647298276424408, + 0.014989792369306087, + 0.05138033255934715, + 0.0164466742426157, + 0.018834279850125313, + 0.058933671563863754, + 0.01281164400279522, + 0.021229004487395287, + 0.019676877185702324, + 0.012377269566059113, + 0.012856808491051197, + 0.006092493422329426, + 0.007959136739373207, + 0.008006393909454346, + 0.007187931332737207, + 0.008969385176897049, + 0.004373311530798674, + 0.008146402426064014, + 0.010220942087471485, + 0.010827898979187012, + 0.008178243413567543, + 0.0064870142377913, + 0.013737007975578308, + 0.009693387895822525, + 0.005453519523143768, + 0.007395016495138407, + 0.017955400049686432, + 0.01642424799501896, + 0.021122073754668236, + 0.01802467368543148, + 0.019149407744407654, + 0.014593126252293587, + 0.015112466178834438, + 0.020415008068084717, + 0.009841060265898705, + 0.014841687865555286, + 0.011360401287674904, + 0.013081403449177742, + 0.012003879994153976, + 0.009672936983406544, + 0.011381465941667557, + 0.017588896676898003, + 0.014544712379574776, + 0.018092088401317596, + 0.010953300632536411, + 0.01083030179142952, + 0.020600153133273125, + 0.012234730646014214, + 0.01976054348051548, + 0.02374756895005703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.0006349454633891582 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_tensor([6584, 6584])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_tensor([6584, 6584])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a07a9c579db325291a4a9f31a3af0e53f95c6f2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_tensor([6584, 6584])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up, angled slightly downward, begins a slow and steady zoom, focusing on a pair of hands meticulously placing books. With long fingers and neatly manicured nails, the hands carefully guide a small stack of antique leather-bound volumes—one a deep burgundy, the other a forest green, their spines adorned with faded gold leaf filigree. They are slid into the last empty space on a dark, polished mahogany bookshelf already crowded with similar books. The movement is deliberate and gentle, the books gliding smoothly across the wood until they are perfectly seated, the final effect soft and satisfying. The entire scene is bathed in warm, low-key light from an off-camera source, casting a soft golden glow that accentuates the rich texture of the leather and the deep grain of the wood, creating an intimate and serene quiet with a distinctly warm, cinematic feel.", + "seed": 12336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_6584_seed_12336_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027850631740875542, + "vlm_logprob": -0.027850631740875542, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.027850631740875542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010848354548215866 + ], + "opposite_logprob_tokens": [ + -4.885848522186279 + ], + "logprob_reward": -0.010848354548215866, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010848354548215866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1573040783405304, + 0.0586373507976532, + 0.06774210929870605, + 0.186569944024086, + 0.09517981112003326, + 0.0666436031460762, + 0.1036791279911995, + 0.15825265645980835, + 0.08103004097938538, + 0.024961311370134354 + ], + "credit_map": [ + 0.019112998619675636, + 0.02717292308807373, + 0.020463990047574043, + 0.016561001539230347, + 0.021512195467948914, + 0.02037900499999523, + 0.018195979297161102, + 0.013905981555581093, + 0.0052073863334953785, + 0.006066679488867521, + 0.008219543844461441, + 0.006924711633473635, + 0.007504691835492849, + 0.0071781352162361145, + 0.006413621362298727, + 0.011122580617666245, + 0.005880721379071474, + 0.009982384741306305, + 0.006044940557330847, + 0.005940400063991547, + 0.014247437007725239, + 0.011694710701704025, + 0.007271850015968084, + 0.006679666228592396, + 0.016120539978146553, + 0.01136376429349184, + 0.050299081951379776, + 0.024044500663876534, + 0.03377555310726166, + 0.020362600684165955, + 0.011650916188955307, + 0.0189529899507761, + 0.008851619437336922, + 0.007256544195115566, + 0.022217169404029846, + 0.009183072485029697, + 0.012197618372738361, + 0.011035824194550514, + 0.00999116338789463, + 0.014446799643337727, + 0.01040561217814684, + 0.006602209992706776, + 0.010306823067367077, + 0.00838456116616726, + 0.00775314774364233, + 0.007798478007316589, + 0.007792945019900799, + 0.007599824573844671, + 0.011093365028500557, + 0.0072565521113574505, + 0.014643537811934948, + 0.01663900725543499, + 0.015774162486195564, + 0.01577788032591343, + 0.012250312604010105, + 0.010244321078062057, + 0.014815525151789188, + 0.014216182753443718, + 0.034074123948812485, + 0.02534530684351921, + 0.020642558112740517, + 0.014986134134232998, + 0.018415680155158043, + 0.015757139772176743, + 0.009689008817076683, + 0.010632074438035488, + 0.012512749060988426, + 0.010370735079050064, + 0.01155156921595335, + 0.007932051084935665, + 0.007296552881598473, + 0.011045293882489204, + 0.003328298684209585, + 0.0028489308897405863, + 0.004062863532453775, + 0.002583653898909688, + 0.0031466949731111526, + 0.003358434420078993, + 0.0033441148698329926, + 0.002288318704813719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.010848354548215866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014229030348360538 + ], + "opposite_logprob_tokens": [ + -4.264228820800781 + ], + "logprob_reward": -0.014229030348360538, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014229030348360538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3424406349658966, + 0.10418302565813065, + 0.14087508618831635, + 0.08365745842456818, + 0.05216711014509201, + 0.06180558726191521, + 0.02940838225185871, + 0.02419484220445156, + 0.1263706088066101, + 0.034897271543741226 + ], + "credit_map": [ + 0.031787529587745667, + 0.04971305653452873, + 0.042153485119342804, + 0.045142728835344315, + 0.050635501742362976, + 0.04713420942425728, + 0.047413259744644165, + 0.02846086211502552, + 0.013910275883972645, + 0.016096414998173714, + 0.007657884154468775, + 0.007697986904531717, + 0.015715597197413445, + 0.015351471491158009, + 0.011100593954324722, + 0.016652807593345642, + 0.008434002287685871, + 0.022175129503011703, + 0.016627663746476173, + 0.01047365553677082, + 0.03310193121433258, + 0.027729135006666183, + 0.012493474408984184, + 0.009840080514550209, + 0.0037145279347896576, + 0.005522898864001036, + 0.014823032543063164, + 0.005513189826160669, + 0.03071705624461174, + 0.017308056354522705, + 0.002597315004095435, + 0.0034613851457834244, + 0.008048535324633121, + 0.0035469625145196915, + 0.007456506136804819, + 0.006072496064007282, + 0.008787840604782104, + 0.007896040566265583, + 0.004178959876298904, + 0.00617977324873209, + 0.010671090334653854, + 0.004964631050825119, + 0.011613423004746437, + 0.007174320053309202, + 0.006172950845211744, + 0.006878246087580919, + 0.00631602993234992, + 0.008014895021915436, + 0.0029324758797883987, + 0.0020895833149552345, + 0.0053915223106741905, + 0.0054343510419130325, + 0.0034175666514784098, + 0.004620336927473545, + 0.003021690296009183, + 0.002500856528058648, + 0.002614599419757724, + 0.002823582384735346, + 0.003804740495979786, + 0.003824845189228654, + 0.0029955666977912188, + 0.0019195199711248279, + 0.0031814815010875463, + 0.0030305080581456423, + 0.009713518433272839, + 0.03550603240728378, + 0.03307092934846878, + 0.01170337200164795, + 0.01089214626699686, + 0.00914666149765253, + 0.007122701965272427, + 0.00921524316072464, + 0.0034795457031577826, + 0.0034096743911504745, + 0.0035030159633606672, + 0.00409328518435359, + 0.005692208185791969, + 0.00569663941860199, + 0.004258275032043457, + 0.004764627665281296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.014229030348360538 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033701904118061066 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.033701904118061066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701904118061066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23520585894584656, + 0.08683682233095169, + 0.07115016132593155, + 0.1718941181898117, + 0.0864683985710144, + 0.043027762323617935, + 0.032313648611307144, + 0.03304377943277359, + 0.11176376044750214, + 0.1282956749200821 + ], + "credit_map": [ + 0.018413623794913292, + 0.04099179059267044, + 0.03678586333990097, + 0.030871976166963577, + 0.034292638301849365, + 0.025027552619576454, + 0.03219669312238693, + 0.016625724732875824, + 0.007266640663146973, + 0.010858893394470215, + 0.014164101332426071, + 0.009597500786185265, + 0.014980953186750412, + 0.011167571879923344, + 0.010731371119618416, + 0.0080697787925601, + 0.007135001942515373, + 0.012217806652188301, + 0.008172735571861267, + 0.005717741325497627, + 0.009173452854156494, + 0.009063159115612507, + 0.0075398520566523075, + 0.012130401097238064, + 0.01595720835030079, + 0.01350926049053669, + 0.032877832651138306, + 0.016937309876084328, + 0.0447625108063221, + 0.02754414826631546, + 0.011557127349078655, + 0.008748706430196762, + 0.015680810436606407, + 0.00587495556101203, + 0.016810962930321693, + 0.007461232598870993, + 0.00751313753426075, + 0.011089451611042023, + 0.0064940862357616425, + 0.015543760731816292, + 0.007941986434161663, + 0.0037379758432507515, + 0.0037862255703657866, + 0.005795134697109461, + 0.006106381304562092, + 0.006878006737679243, + 0.004066505469381809, + 0.004715545568615198, + 0.002224154770374298, + 0.002109281253069639, + 0.009178356267511845, + 0.009090408682823181, + 0.002852646866813302, + 0.0021097881253808737, + 0.002418805379420519, + 0.0023302047047764063, + 0.0035758602898567915, + 0.0037537121679633856, + 0.005428723059594631, + 0.005650142207741737, + 0.004673493094742298, + 0.003086384618654847, + 0.003189915558323264, + 0.0036855468060821295, + 0.00609604874625802, + 0.0346883125603199, + 0.03120860643684864, + 0.009617563337087631, + 0.009645684622228146, + 0.00830142293125391, + 0.005360242445021868, + 0.0068458677269518375, + 0.008524373173713684, + 0.01202237606048584, + 0.03243261203169823, + 0.017098836600780487, + 0.010367116890847683, + 0.025444798171520233, + 0.012462565675377846, + 0.009942987933754921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.033701904118061066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015147175872698426 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015147175872698426, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015147175872698426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16633494198322296, + 0.15604346990585327, + 0.04807671532034874, + 0.06822752952575684, + 0.03717100992798805, + 0.03643552586436272, + 0.11412469297647476, + 0.10089476406574249, + 0.2306164801120758, + 0.0420747846364975 + ], + "credit_map": [ + 0.016562974080443382, + 0.020374706014990807, + 0.023972898721694946, + 0.0198126882314682, + 0.022488737478852272, + 0.02182537131011486, + 0.020746445283293724, + 0.020551130175590515, + 0.010941620916128159, + 0.014872187748551369, + 0.020611902698874474, + 0.014046584255993366, + 0.037350915372371674, + 0.02083892934024334, + 0.016582833603024483, + 0.02079850062727928, + 0.0051942975260317326, + 0.007737469393759966, + 0.00609857402741909, + 0.004791093058884144, + 0.005997045896947384, + 0.0066455090418457985, + 0.005143135320395231, + 0.006469591520726681, + 0.008634116500616074, + 0.006720241624861956, + 0.00866603571921587, + 0.007170351687818766, + 0.011972896754741669, + 0.009516891092061996, + 0.007180186919867992, + 0.00836680643260479, + 0.00393323739990592, + 0.002610183088108897, + 0.007770034018903971, + 0.0029281130991876125, + 0.003558171447366476, + 0.005224865395575762, + 0.0034070967230945826, + 0.007739309687167406, + 0.0037430264055728912, + 0.003463364439085126, + 0.0035525797866284847, + 0.006701789330691099, + 0.0028967969119548798, + 0.004145980346947908, + 0.005328241270035505, + 0.006603749934583902, + 0.01662128046154976, + 0.009008373133838177, + 0.013545135036110878, + 0.013934966176748276, + 0.01892404444515705, + 0.01893080770969391, + 0.010733329690992832, + 0.012426754459738731, + 0.014297224581241608, + 0.009196674451231956, + 0.014484035782516003, + 0.016129426658153534, + 0.012378468178212643, + 0.010298366658389568, + 0.01174391433596611, + 0.012366652488708496, + 0.031307775527238846, + 0.05460308492183685, + 0.0582934133708477, + 0.017299523577094078, + 0.020564155653119087, + 0.01886054314672947, + 0.013557597994804382, + 0.016130397096276283, + 0.003201900515705347, + 0.0037187314592301846, + 0.009464421309530735, + 0.0041906037367880344, + 0.0031940885819494724, + 0.007618575822561979, + 0.007056402042508125, + 0.003630060702562332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0015147175872698426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789591521024704 + ], + "opposite_logprob_tokens": [ + -2.5789592266082764 + ], + "logprob_reward": -0.0789591521024704, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789591521024704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12738114595413208, + 0.20547351241111755, + 0.044947315007448196, + 0.22791941463947296, + 0.042138051241636276, + 0.05067851021885872, + 0.06160376965999603, + 0.08073634654283524, + 0.062457576394081116, + 0.09666439890861511 + ], + "credit_map": [ + 0.015642084181308746, + 0.017380809411406517, + 0.025675896555185318, + 0.016947807744145393, + 0.015074850991368294, + 0.013976233080029488, + 0.01116057951003313, + 0.011522887274622917, + 0.014824645593762398, + 0.020204856991767883, + 0.035275738686323166, + 0.016127387061715126, + 0.06341677904129028, + 0.02853839285671711, + 0.016521234065294266, + 0.010564451105892658, + 0.004309166222810745, + 0.008817438036203384, + 0.004703047685325146, + 0.004017225932329893, + 0.007733490783721209, + 0.006275998894125223, + 0.00448466744273901, + 0.0046062772162258625, + 0.026687826961278915, + 0.0234444011002779, + 0.04183980077505112, + 0.027363980188965797, + 0.0365648977458477, + 0.02968653291463852, + 0.02010486088693142, + 0.0222270917147398, + 0.004668956622481346, + 0.0038560039829462767, + 0.003924877382814884, + 0.004659533500671387, + 0.008955132216215134, + 0.007704118732362986, + 0.004708335269242525, + 0.0036610933020710945, + 0.00615885853767395, + 0.005170357879251242, + 0.006869582459330559, + 0.006511108949780464, + 0.00845840573310852, + 0.0073835477232933044, + 0.005475666373968124, + 0.004650978837162256, + 0.005594378337264061, + 0.002903634449467063, + 0.015614638105034828, + 0.017756538465619087, + 0.005252655595541, + 0.005021292716264725, + 0.004755208734422922, + 0.004705417435616255, + 0.009251506067812443, + 0.007200576364994049, + 0.015498108230531216, + 0.015298291109502316, + 0.006023300811648369, + 0.006718712393194437, + 0.010933252051472664, + 0.009812593460083008, + 0.011256763711571693, + 0.011191967874765396, + 0.010402215644717216, + 0.0055537838488817215, + 0.007084870710968971, + 0.006694508250802755, + 0.004508323967456818, + 0.005765139125287533, + 0.007194434758275747, + 0.00929301232099533, + 0.02430356852710247, + 0.011231914162635803, + 0.007258772850036621, + 0.01808391511440277, + 0.01223597303032875, + 0.007062800694257021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0789591521024704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, angled slightly downward, begins a slow and steady zoom, focusing on a pair of hands meticulously placing books. With long fingers and neatly manicured nails, the hands carefully guide a small stack of antique leather-bound volumes—one a deep burgundy, the other a forest green, their spines adorned with faded gold leaf filigree. They are slid into the last empty space on a dark, polished mahogany bookshelf already crowded with similar books. The movement is deliberate and gentle, the books gliding smoothly across the wood until they are perfectly seated, the final effect soft and satisfying. The entire scene is bathed in warm, low-key light from an off-camera source, casting a soft golden glow that accentuates the rich texture of the leather and the deep grain of the wood, creating an intimate and serene quiet with a distinctly warm, cinematic feel.", + "seed": 12337, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_6584_seed_12337_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020622534584254025, + "vlm_logprob": -0.020622534584254025, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.020622534584254025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01704695075750351 + ], + "opposite_logprob_tokens": [ + -4.267046928405762 + ], + "logprob_reward": -0.01704695075750351, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01704695075750351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10355103015899658, + 0.0878981500864029, + 0.08665179461240768, + 0.0572255477309227, + 0.06250138580799103, + 0.046304911375045776, + 0.04733694717288017, + 0.10950999706983566, + 0.1984098255634308, + 0.2006104737520218 + ], + "credit_map": [ + 0.007179037667810917, + 0.019445080310106277, + 0.011165803298354149, + 0.018679670989513397, + 0.00823761336505413, + 0.013042056933045387, + 0.015318945981562138, + 0.010482825338840485, + 0.00904663372784853, + 0.014546520076692104, + 0.012189095839858055, + 0.011676428839564323, + 0.0066580031998455524, + 0.0061882538720965385, + 0.005207967013120651, + 0.02238525077700615, + 0.009863216429948807, + 0.010287221521139145, + 0.01629035733640194, + 0.011768030934035778, + 0.011180142872035503, + 0.008100743405520916, + 0.009311622008681297, + 0.009850461035966873, + 0.008535386063158512, + 0.006805798038840294, + 0.016003893688321114, + 0.005026660393923521, + 0.0036183586344122887, + 0.004103655461221933, + 0.005960908718407154, + 0.007170885801315308, + 0.008765406906604767, + 0.006794705521315336, + 0.008714194409549236, + 0.006887872703373432, + 0.007760625798255205, + 0.0060933842323720455, + 0.009011607617139816, + 0.008473586291074753, + 0.008916343562304974, + 0.005931224673986435, + 0.007737127598375082, + 0.005940965376794338, + 0.0031589139252901077, + 0.00554559426382184, + 0.004159789532423019, + 0.004914957098662853, + 0.007376607041805983, + 0.0067288316786289215, + 0.008189945481717587, + 0.006612828932702541, + 0.006255751010030508, + 0.004531201906502247, + 0.0042335563339293, + 0.0034082240890711546, + 0.010536552406847477, + 0.016513502225279808, + 0.01641075871884823, + 0.016306625679135323, + 0.013853632844984531, + 0.01176578737795353, + 0.013214261271059513, + 0.010908872820436954, + 0.021438291296362877, + 0.033922530710697174, + 0.027097349986433983, + 0.018218200653791428, + 0.02934807538986206, + 0.02360592968761921, + 0.02404201775789261, + 0.020737428218126297, + 0.033490732312202454, + 0.013879864476621151, + 0.07538711279630661, + 0.0474223829805851, + 0.004787523299455643, + 0.005555126816034317, + 0.00995579268783331, + 0.01013194490224123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.01704695075750351 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610848680138588 + ], + "opposite_logprob_tokens": [ + -4.141108512878418 + ], + "logprob_reward": -0.01610848680138588, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610848680138588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10765916109085083, + 0.1365070790052414, + 0.0437457337975502, + 0.033034585416316986, + 0.02808358147740364, + 0.03760194778442383, + 0.028132596984505653, + 0.07186784595251083, + 0.05789388343691826, + 0.45547357201576233 + ], + "credit_map": [ + 0.007358686998486519, + 0.014056987129151821, + 0.010824724100530148, + 0.016190042719244957, + 0.014927309937775135, + 0.011339958757162094, + 0.01118397992104292, + 0.021777469664812088, + 0.014727587811648846, + 0.016814742237329483, + 0.012944474816322327, + 0.02371710166335106, + 0.014880391769111156, + 0.022146150469779968, + 0.011265932582318783, + 0.020010685548186302, + 0.005821113474667072, + 0.0047059981152415276, + 0.00673644756898284, + 0.006100138183683157, + 0.005995863117277622, + 0.006142199970781803, + 0.0033602523617446423, + 0.004883721005171537, + 0.004375901538878679, + 0.0030474013183265924, + 0.008645225316286087, + 0.0035794537980109453, + 0.003171969437971711, + 0.0024826826993376017, + 0.0023785834200680256, + 0.005353367887437344, + 0.004424407612532377, + 0.0035493646282702684, + 0.0034501429181545973, + 0.0033219719771295786, + 0.0026223899330943823, + 0.003620038041844964, + 0.003477389458566904, + 0.003617875976487994, + 0.008122353814542294, + 0.003823360428214073, + 0.005310251843184233, + 0.003899879287928343, + 0.004452875815331936, + 0.0037796974647790194, + 0.004020834341645241, + 0.004192695952951908, + 0.008046041242778301, + 0.0030107300262898207, + 0.004740905482321978, + 0.0022158611100167036, + 0.002875660778954625, + 0.0023822749499231577, + 0.0025549111887812614, + 0.0023062126711010933, + 0.011953018605709076, + 0.011033318936824799, + 0.009935895912349224, + 0.01043110154569149, + 0.007876269519329071, + 0.005193754099309444, + 0.008595895022153854, + 0.006848589517176151, + 0.009956320747733116, + 0.009093556553125381, + 0.008001359179615974, + 0.006295128259807825, + 0.006719987373799086, + 0.00735750887542963, + 0.005325441714376211, + 0.005144580267369747, + 0.0451924167573452, + 0.0266775693744421, + 0.16157519817352295, + 0.14784622192382812, + 0.018488330766558647, + 0.012286709621548653, + 0.026906471699476242, + 0.01650066487491131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01610848680138588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2221376597881317, + 0.2371288239955902, + 0.058215755969285965, + 0.10560479015111923, + 0.03288125619292259, + 0.051755212247371674, + 0.022598996758461, + 0.0455702506005764, + 0.03498585894703865, + 0.18912144005298615 + ], + "credit_map": [ + 0.02036099322140217, + 0.036658789962530136, + 0.020824851468205452, + 0.031051263213157654, + 0.025845790281891823, + 0.0248391255736351, + 0.028577957302331924, + 0.03397887945175171, + 0.02480676956474781, + 0.027284305542707443, + 0.030110130086541176, + 0.04385462403297424, + 0.024455353617668152, + 0.032486479729413986, + 0.019088316708803177, + 0.03504283353686333, + 0.01120943482965231, + 0.0076757799834012985, + 0.009033776819705963, + 0.009185826405882835, + 0.005060781259089708, + 0.003970249090343714, + 0.00420429278165102, + 0.007875613868236542, + 0.015874676406383514, + 0.008876628242433071, + 0.03159194067120552, + 0.0056346263736486435, + 0.0069814217276871204, + 0.004902461543679237, + 0.005986778996884823, + 0.02575625292956829, + 0.007431374862790108, + 0.004553521052002907, + 0.005272297654300928, + 0.003771221963688731, + 0.0033348016440868378, + 0.0026428191922605038, + 0.00313573912717402, + 0.002739479998126626, + 0.007527115289121866, + 0.00699627585709095, + 0.006127423606812954, + 0.006602487992495298, + 0.006678874138742685, + 0.005576497875154018, + 0.006158627104014158, + 0.00608791084960103, + 0.003315457608550787, + 0.0027873965445905924, + 0.003143347566947341, + 0.0035947049036622047, + 0.0026093032211065292, + 0.002044866792857647, + 0.002904040738940239, + 0.0021998793818056583, + 0.011061957105994225, + 0.005806474015116692, + 0.006652150768786669, + 0.007829477079212666, + 0.00360760441981256, + 0.0028255549259483814, + 0.00454748747870326, + 0.003239545039832592, + 0.005383520387113094, + 0.004919803701341152, + 0.004214268643409014, + 0.0039013836067169905, + 0.004830228164792061, + 0.0042488169856369495, + 0.004193408880382776, + 0.003294427879154682, + 0.019637923687696457, + 0.016642441973090172, + 0.0480838380753994, + 0.06371700018644333, + 0.0063499039970338345, + 0.00804354902356863, + 0.012143641710281372, + 0.014503146521747112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003618124406784773 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.003618124406784773, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003618124406784773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3357020616531372, + 0.11091281473636627, + 0.09093032032251358, + 0.10727136582136154, + 0.03241398185491562, + 0.04287194460630417, + 0.02931058034300804, + 0.06221647188067436, + 0.09344904124736786, + 0.09492143988609314 + ], + "credit_map": [ + 0.038077495992183685, + 0.056853462010622025, + 0.04572969675064087, + 0.03984550014138222, + 0.029519252479076385, + 0.0409945473074913, + 0.028308121487498283, + 0.056373972445726395, + 0.011731564998626709, + 0.017125040292739868, + 0.020550504326820374, + 0.015188408084213734, + 0.011608299799263477, + 0.012004300951957703, + 0.009742995724081993, + 0.012961703352630138, + 0.008746153675019741, + 0.009724321775138378, + 0.01542665995657444, + 0.014423567801713943, + 0.012254002504050732, + 0.010713460855185986, + 0.009778640232980251, + 0.00986351165920496, + 0.014506828971207142, + 0.010110764764249325, + 0.03303121030330658, + 0.008011681027710438, + 0.004756264854222536, + 0.005960836540907621, + 0.00576640572398901, + 0.02512737549841404, + 0.00546211376786232, + 0.00407750066369772, + 0.004171145148575306, + 0.003164032706990838, + 0.003404913702979684, + 0.003623819909989834, + 0.0038815566804260015, + 0.004628898110240698, + 0.006881559267640114, + 0.005470470525324345, + 0.005344895180314779, + 0.005958327557891607, + 0.00439027976244688, + 0.004356396850198507, + 0.0051843696273863316, + 0.0052856458351016045, + 0.008058284409344196, + 0.004131503868848085, + 0.005124678369611502, + 0.0026823028456419706, + 0.0026741353794932365, + 0.0022031336557120085, + 0.002055814955383539, + 0.002380726858973503, + 0.010188811458647251, + 0.006284607108682394, + 0.007852966897189617, + 0.008267279714345932, + 0.008058092556893826, + 0.006929726805537939, + 0.005992781836539507, + 0.008642205968499184, + 0.01231428049504757, + 0.013734121806919575, + 0.012057142332196236, + 0.0102727971971035, + 0.011852176859974861, + 0.013388893567025661, + 0.009784192778170109, + 0.010045438073575497, + 0.015254444442689419, + 0.012918542139232159, + 0.019311336800456047, + 0.019349338486790657, + 0.00766116613522172, + 0.0063287182711064816, + 0.008400184102356434, + 0.005697703920304775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003618124406784773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043046727776527405 + ], + "opposite_logprob_tokens": [ + -3.168046712875366 + ], + "logprob_reward": -0.043046727776527405, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043046727776527405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2544691860675812, + 0.189938023686409, + 0.14019155502319336, + 0.053047969937324524, + 0.04171418398618698, + 0.034223590046167374, + 0.08724567294120789, + 0.04615490883588791, + 0.06266171485185623, + 0.09035316854715347 + ], + "credit_map": [ + 0.028245529159903526, + 0.03438590094447136, + 0.03437435254454613, + 0.031582582741975784, + 0.02863945998251438, + 0.031135110184550285, + 0.019858496263623238, + 0.04624775052070618, + 0.01881558820605278, + 0.0185577142983675, + 0.0292576402425766, + 0.02863384038209915, + 0.02369312010705471, + 0.02902304194867611, + 0.017126619815826416, + 0.024830462411046028, + 0.01268745493143797, + 0.013866141438484192, + 0.029685402289032936, + 0.022362206131219864, + 0.019577544182538986, + 0.013430263847112656, + 0.012696615420281887, + 0.01588592864573002, + 0.007460160180926323, + 0.005868151783943176, + 0.012591925449669361, + 0.007687934674322605, + 0.0034564179368317127, + 0.004213217180222273, + 0.004228389356285334, + 0.00754177151247859, + 0.007809144910424948, + 0.0046636308543384075, + 0.006937968544661999, + 0.006394562777131796, + 0.003671051934361458, + 0.0030951728112995625, + 0.0048531885258853436, + 0.004289461765438318, + 0.0051776934415102005, + 0.004873973783105612, + 0.005509223323315382, + 0.004724869970232248, + 0.003205161076039076, + 0.003986213821917772, + 0.0031360296998173, + 0.0036104258615523577, + 0.015616004355251789, + 0.009970362298190594, + 0.022553041577339172, + 0.010606455616652966, + 0.008171183988451958, + 0.005497531034052372, + 0.006762141361832619, + 0.008068956434726715, + 0.005803512874990702, + 0.010147850960493088, + 0.005840687546879053, + 0.005149775184690952, + 0.005650385282933712, + 0.00443638488650322, + 0.00442470982670784, + 0.004701601807028055, + 0.008372204378247261, + 0.010919512249529362, + 0.008524914272129536, + 0.007228733506053686, + 0.00760642671957612, + 0.0079051423817873, + 0.006258136592805386, + 0.005846642889082432, + 0.010742518119513988, + 0.010055633261799812, + 0.026866519823670387, + 0.01979147456586361, + 0.005351093597710133, + 0.00492911646142602, + 0.007310923654586077, + 0.005305899307131767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043046727776527405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, angled slightly downward, begins a slow and steady zoom, focusing on a pair of hands meticulously placing books. With long fingers and neatly manicured nails, the hands carefully guide a small stack of antique leather-bound volumes—one a deep burgundy, the other a forest green, their spines adorned with faded gold leaf filigree. They are slid into the last empty space on a dark, polished mahogany bookshelf already crowded with similar books. The movement is deliberate and gentle, the books gliding smoothly across the wood until they are perfectly seated, the final effect soft and satisfying. The entire scene is bathed in warm, low-key light from an off-camera source, casting a soft golden glow that accentuates the rich texture of the leather and the deep grain of the wood, creating an intimate and serene quiet with a distinctly warm, cinematic feel.", + "seed": 12338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_6584_seed_12338_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021229078294709325, + "vlm_logprob": -0.021229078294709325, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.021229078294709325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014721221290528774 + ], + "opposite_logprob_tokens": [ + -4.514721393585205 + ], + "logprob_reward": -0.014721221290528774, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014721221290528774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920611500740051, + 0.06214664503931999, + 0.04427317529916763, + 0.07596542686223984, + 0.020554987713694572, + 0.04478142410516739, + 0.07317298650741577, + 0.02782498486340046, + 0.3050244152545929, + 0.05419478192925453 + ], + "credit_map": [ + 0.02161291427910328, + 0.0853428766131401, + 0.021066583693027496, + 0.012134164571762085, + 0.024606017395853996, + 0.09539424628019333, + 0.01775766722857952, + 0.014146645553410053, + 0.0047769080847501755, + 0.006648530252277851, + 0.006701395846903324, + 0.007817941717803478, + 0.009299607947468758, + 0.011619266122579575, + 0.007441437337547541, + 0.007841560058295727, + 0.004331487696617842, + 0.004519185051321983, + 0.006736350245773792, + 0.00614467216655612, + 0.005626120138913393, + 0.006824246142059565, + 0.004900986794382334, + 0.005190127994865179, + 0.009323511272668839, + 0.009717588312923908, + 0.010681776329874992, + 0.011133592575788498, + 0.0072415233589708805, + 0.009415462613105774, + 0.008805694989860058, + 0.009646275080740452, + 0.0028797893319278955, + 0.0018223157385364175, + 0.0026309939566999674, + 0.003113812068477273, + 0.0019617704674601555, + 0.0027605807408690453, + 0.0028117792680859566, + 0.002573945326730609, + 0.004639971535652876, + 0.003329360159114003, + 0.00315540237352252, + 0.004991995170712471, + 0.004554522689431906, + 0.009592393413186073, + 0.006169842556118965, + 0.00834793783724308, + 0.0036026237066835165, + 0.003680676920339465, + 0.00453453604131937, + 0.004398160148411989, + 0.0047580343671143055, + 0.025178784504532814, + 0.006159474141895771, + 0.020860694348812103, + 0.002839961089193821, + 0.0021674376912415028, + 0.003764940658584237, + 0.0035699675790965557, + 0.004147735890001059, + 0.004738318733870983, + 0.0038473261520266533, + 0.0027492973022162914, + 0.02050764672458172, + 0.007611033972352743, + 0.12085875123739243, + 0.012793042697012424, + 0.018385857343673706, + 0.01931249164044857, + 0.020418189465999603, + 0.08513742685317993, + 0.003093863371759653, + 0.0027897951658815145, + 0.0054607512429356575, + 0.0033810948953032494, + 0.0033944614697247744, + 0.014014636166393757, + 0.006547497119754553, + 0.015512688085436821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.014721221290528774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610168255865574 + ], + "opposite_logprob_tokens": [ + -4.141101837158203 + ], + "logprob_reward": -0.01610168255865574, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610168255865574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48098087310791016, + 0.024242162704467773, + 0.022298773750662804, + 0.03902308642864227, + 0.01953466609120369, + 0.040193960070610046, + 0.026497334241867065, + 0.031808916479349136, + 0.04640175774693489, + 0.2690184414386749 + ], + "credit_map": [ + 0.020221561193466187, + 0.17573025822639465, + 0.014297557063400745, + 0.012224490754306316, + 0.03846638277173042, + 0.17418353259563446, + 0.02886638604104519, + 0.016990749165415764, + 0.002765369601547718, + 0.002967770444229245, + 0.00234990194439888, + 0.0023287066724151373, + 0.0038926214911043644, + 0.004721073433756828, + 0.002615354722365737, + 0.0026013669557869434, + 0.0025693734642118216, + 0.0023339681793004274, + 0.002002087188884616, + 0.0017374188173562288, + 0.0039000350516289473, + 0.0043512615375220776, + 0.0016774754039943218, + 0.003727156203240156, + 0.004539874382317066, + 0.0040326532907783985, + 0.003090860554948449, + 0.003462023800238967, + 0.005188556853681803, + 0.009080213494598866, + 0.004600656218826771, + 0.005028247367590666, + 0.0021588685922324657, + 0.0016775182448327541, + 0.0026559727266430855, + 0.002229753416031599, + 0.0022647345904260874, + 0.00343178934417665, + 0.0027275672182440758, + 0.0023884635884314775, + 0.00532563729211688, + 0.0025758554693311453, + 0.0032434596214443445, + 0.0030745049007236958, + 0.004678172059357166, + 0.008530822582542896, + 0.005248921923339367, + 0.007516589481383562, + 0.0026422045193612576, + 0.0012747615110129118, + 0.002493735635653138, + 0.002621556632220745, + 0.0034832011442631483, + 0.006434951443225145, + 0.002651902148500085, + 0.004895024001598358, + 0.0028145865071564913, + 0.0035839849151670933, + 0.0038589397445321083, + 0.004236817825585604, + 0.005828965920954943, + 0.003873424604535103, + 0.004387824796140194, + 0.003224375192075968, + 0.003033326705917716, + 0.0016430887626484036, + 0.015604443848133087, + 0.002401153789833188, + 0.0035146705340594053, + 0.004905499052256346, + 0.0036254909355193377, + 0.01167408935725689, + 0.013735402375459671, + 0.009556234814226627, + 0.04289380833506584, + 0.014733449555933475, + 0.015963613986968994, + 0.09225380420684814, + 0.027004340663552284, + 0.05287780985236168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01610168255865574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3610902726650238, + 0.031738851219415665, + 0.03323211520910263, + 0.05437850579619408, + 0.011528478935360909, + 0.009129127487540245, + 0.0870799869298935, + 0.04276531934738159, + 0.25273212790489197, + 0.1163252741098404 + ], + "credit_map": [ + 0.020671052858233452, + 0.12357169389724731, + 0.014844021759927273, + 0.01067901961505413, + 0.028127336874604225, + 0.12445173412561417, + 0.021200761198997498, + 0.01754463091492653, + 0.0027698318008333445, + 0.003462729509919882, + 0.0021392423659563065, + 0.00234697456471622, + 0.005373424384742975, + 0.006827360950410366, + 0.005389605648815632, + 0.003429680597037077, + 0.003730477998033166, + 0.002604616107419133, + 0.0033800387755036354, + 0.0031784421298652887, + 0.004216640722006559, + 0.005370309576392174, + 0.0039767189882695675, + 0.0067748730070889, + 0.004790831822901964, + 0.004914773162454367, + 0.0064032962545752525, + 0.005823694635182619, + 0.007744764443486929, + 0.010151353664696217, + 0.006436172407120466, + 0.008113618940114975, + 0.001340453396551311, + 0.0010182879632338881, + 0.0015491890953853726, + 0.0012562454212456942, + 0.0014775838935747743, + 0.0016712992219254375, + 0.0017580270068719983, + 0.0014573921216651797, + 0.00098138814792037, + 0.0005270242691040039, + 0.000802849477622658, + 0.0008510990883223712, + 0.0010158794466406107, + 0.0021182282362133265, + 0.0011043593985959888, + 0.0017282988410443068, + 0.004488620441406965, + 0.0040404158644378185, + 0.00466002244502306, + 0.003845736850053072, + 0.006699422840029001, + 0.028945140540599823, + 0.008697455748915672, + 0.025703176856040955, + 0.00334720965474844, + 0.004547920543700457, + 0.004860007204115391, + 0.005897135473787785, + 0.007278824225068092, + 0.0062523772940039635, + 0.0062436130829155445, + 0.004338233266025782, + 0.01718575693666935, + 0.00588566018268466, + 0.10198848694562912, + 0.01036585122346878, + 0.01462374348193407, + 0.01453079842031002, + 0.016692381352186203, + 0.07145943492650986, + 0.007207673974335194, + 0.004682138096541166, + 0.008343563415110111, + 0.0035852014552801847, + 0.004479680676013231, + 0.039528459310531616, + 0.010480417869985104, + 0.03801814466714859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021975908894091845 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021975908894091845, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021975908894091845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32451707124710083, + 0.09479241818189621, + 0.02821403369307518, + 0.04956653341650963, + 0.07608361542224884, + 0.055170875042676926, + 0.02944432944059372, + 0.12422040104866028, + 0.1392427235841751, + 0.07874801009893417 + ], + "credit_map": [ + 0.015084093436598778, + 0.10817164927721024, + 0.022438928484916687, + 0.011943548917770386, + 0.028390565887093544, + 0.10329627245664597, + 0.021790094673633575, + 0.01340192649513483, + 0.011060019955039024, + 0.009862075559794903, + 0.009902620688080788, + 0.009867463260889053, + 0.016073282808065414, + 0.012964580208063126, + 0.016723757609725, + 0.008338617160916328, + 0.003893523942679167, + 0.002503075869753957, + 0.003257288131862879, + 0.0033445570152252913, + 0.003296137787401676, + 0.004775576759129763, + 0.002618917031213641, + 0.00452495738863945, + 0.0051870387978851795, + 0.004802026320248842, + 0.007117921020835638, + 0.006439604796469212, + 0.006104830652475357, + 0.006901714019477367, + 0.005290688015520573, + 0.007722705136984587, + 0.008246279321610928, + 0.005113894119858742, + 0.006617567036300898, + 0.007540855556726456, + 0.011428619734942913, + 0.01720624603331089, + 0.009510827250778675, + 0.010419325903058052, + 0.006360722705721855, + 0.006089904345571995, + 0.006059701554477215, + 0.006599241402000189, + 0.007836204953491688, + 0.007217016536742449, + 0.007935091853141785, + 0.007072991691529751, + 0.0023137512616813183, + 0.0017933391500264406, + 0.0023609839845448732, + 0.001902134739793837, + 0.002149313921108842, + 0.008874575607478619, + 0.00314131285995245, + 0.0069089182652533054, + 0.011256321333348751, + 0.013975664973258972, + 0.014239583164453506, + 0.01349398959428072, + 0.02991924248635769, + 0.013853509910404682, + 0.014574510045349598, + 0.012907584197819233, + 0.009378207847476006, + 0.0037785593885928392, + 0.05345860868692398, + 0.006468393374234438, + 0.008605663664638996, + 0.010116799734532833, + 0.009596734307706356, + 0.03783976286649704, + 0.005007782485336065, + 0.0034458793234080076, + 0.00663252966478467, + 0.004018913488835096, + 0.005559565499424934, + 0.025895340368151665, + 0.007287671789526939, + 0.020900318399071693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0021975908894091845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054937709122896194 + ], + "opposite_logprob_tokens": [ + -2.9299376010894775 + ], + "logprob_reward": -0.054937709122896194, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054937709122896194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4154576063156128, + 0.10277571529150009, + 0.021856248378753662, + 0.052532777190208435, + 0.020707767456769943, + 0.039912912994623184, + 0.009483394213020802, + 0.03014003485441208, + 0.2537226974964142, + 0.05341080576181412 + ], + "credit_map": [ + 0.01517160888761282, + 0.15079307556152344, + 0.017045648768544197, + 0.01223673950880766, + 0.03703805431723595, + 0.1406140774488449, + 0.02654523216187954, + 0.016013188287615776, + 0.008580834604799747, + 0.010408516973257065, + 0.006635294761508703, + 0.006974739022552967, + 0.016199924051761627, + 0.024533379822969437, + 0.01616150699555874, + 0.013281515799462795, + 0.003338464768603444, + 0.0019173325272276998, + 0.0019974582828581333, + 0.0030036424286663532, + 0.0029393378645181656, + 0.003241664730012417, + 0.002230178564786911, + 0.003188170026987791, + 0.0035026471596211195, + 0.004746499937027693, + 0.006456200033426285, + 0.0054873195476830006, + 0.00670692278072238, + 0.01058573741465807, + 0.006031876429915428, + 0.009015575982630253, + 0.0022485842928290367, + 0.0016185786807909608, + 0.0017517376691102982, + 0.0017278071027249098, + 0.002774430438876152, + 0.004528177436441183, + 0.0035278433933854103, + 0.0025306078605353832, + 0.005100754555314779, + 0.003587438492104411, + 0.003294621128588915, + 0.004536954220384359, + 0.004458393901586533, + 0.006865240633487701, + 0.00633407523855567, + 0.005735435057431459, + 0.0010951178846880794, + 0.0008494395297020674, + 0.0012603576760739088, + 0.001115750172175467, + 0.0011972186621278524, + 0.0013854242861270905, + 0.0012703579850494862, + 0.0013097279006615281, + 0.0028171211015433073, + 0.003562350058928132, + 0.0037761302664875984, + 0.0028738866094499826, + 0.005971899256110191, + 0.003838067874312401, + 0.004308988805860281, + 0.0029915894847363234, + 0.017062192782759666, + 0.005833607632666826, + 0.10209440439939499, + 0.010604873299598694, + 0.01415715180337429, + 0.014557188376784325, + 0.01659693755209446, + 0.07281634211540222, + 0.00344786886125803, + 0.002426942577585578, + 0.006175021640956402, + 0.002710617845878005, + 0.002058373298496008, + 0.014816767536103725, + 0.005157785955816507, + 0.01661742851138115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.054937709122896194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, angled slightly downward, begins a slow and steady zoom, focusing on a pair of hands meticulously placing books. With long fingers and neatly manicured nails, the hands carefully guide a small stack of antique leather-bound volumes—one a deep burgundy, the other a forest green, their spines adorned with faded gold leaf filigree. They are slid into the last empty space on a dark, polished mahogany bookshelf already crowded with similar books. The movement is deliberate and gentle, the books gliding smoothly across the wood until they are perfectly seated, the final effect soft and satisfying. The entire scene is bathed in warm, low-key light from an off-camera source, casting a soft golden glow that accentuates the rich texture of the leather and the deep grain of the wood, creating an intimate and serene quiet with a distinctly warm, cinematic feel.", + "seed": 12339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_8/index_6584_seed_12339_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025770129542797802, + "vlm_logprob": -0.025770129542797802, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.025770129542797802, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00996342208236456 + ], + "opposite_logprob_tokens": [ + -5.009963512420654 + ], + "logprob_reward": -0.00996342208236456, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00996342208236456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09278654307126999, + 0.03789849206805229, + 0.18918564915657043, + 0.021883081644773483, + 0.02122899517416954, + 0.0746370404958725, + 0.44143837690353394, + 0.06873753666877747, + 0.029976410791277885, + 0.02222788706421852 + ], + "credit_map": [ + 0.0074839238077402115, + 0.01050828117877245, + 0.015074479393661022, + 0.016609735786914825, + 0.007710926700383425, + 0.011756109073758125, + 0.014027527533471584, + 0.009615563787519932, + 0.0031581223011016846, + 0.004062761552631855, + 0.005092952866107225, + 0.007960074581205845, + 0.0024961053859442472, + 0.0047728330828249454, + 0.0034202763345092535, + 0.006935365963727236, + 0.005954356398433447, + 0.0234757661819458, + 0.00469377264380455, + 0.004061032552272081, + 0.004044465254992247, + 0.08255913853645325, + 0.0038302866742014885, + 0.06056682765483856, + 0.0028230866882950068, + 0.0024449038319289684, + 0.0028111974243074656, + 0.0030862565618008375, + 0.002898050006479025, + 0.002593003446236253, + 0.002993285655975342, + 0.0022332987282425165, + 0.0024992560502141714, + 0.0025274395011365414, + 0.0029251123778522015, + 0.002473689615726471, + 0.002061359351500869, + 0.003215199336409569, + 0.002198738045990467, + 0.0033282011281698942, + 0.004171941429376602, + 0.003994021564722061, + 0.006505116820335388, + 0.00537320040166378, + 0.004912860691547394, + 0.020176932215690613, + 0.006573996972292662, + 0.02292897365987301, + 0.020435210317373276, + 0.016997722908854485, + 0.0260581374168396, + 0.024652495980262756, + 0.016842862591147423, + 0.1523098647594452, + 0.02134476602077484, + 0.16279731690883636, + 0.003946562763303518, + 0.0064826770685613155, + 0.00499182241037488, + 0.0034387651830911636, + 0.003320846240967512, + 0.021487858146429062, + 0.003368214936926961, + 0.021700790151953697, + 0.002939664525911212, + 0.004280919674783945, + 0.004373516421765089, + 0.004177061840891838, + 0.0025173663161695004, + 0.005285800434648991, + 0.002323030261322856, + 0.004079049453139305, + 0.001731973490677774, + 0.0021217979956418276, + 0.001866717473603785, + 0.002512147184461355, + 0.0021532804239541292, + 0.0040232655592262745, + 0.0019076494500041008, + 0.0059110564179718494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.00996342208236456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012576202861964703 + ], + "opposite_logprob_tokens": [ + -4.387576103210449 + ], + "logprob_reward": -0.012576202861964703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012576202861964703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07665260881185532, + 0.07457314431667328, + 0.26287728548049927, + 0.02748866006731987, + 0.025163205340504646, + 0.10499849915504456, + 0.1472664475440979, + 0.23906195163726807, + 0.015101718716323376, + 0.02681652456521988 + ], + "credit_map": [ + 0.005263712257146835, + 0.0077459681779146194, + 0.007881543599069118, + 0.013085059821605682, + 0.007872841320931911, + 0.010947463102638721, + 0.008992777206003666, + 0.01486323494464159, + 0.005876682698726654, + 0.008512993343174458, + 0.010135387070477009, + 0.013909731060266495, + 0.004749266430735588, + 0.009471077471971512, + 0.006635024677962065, + 0.015282976441085339, + 0.007588482461869717, + 0.04132051765918732, + 0.007725505158305168, + 0.007666663732379675, + 0.0062510245479643345, + 0.11291662603616714, + 0.005833897273987532, + 0.07357454299926758, + 0.0022893052082508802, + 0.0027603113558143377, + 0.003940374590456486, + 0.004261518828570843, + 0.0026146569289267063, + 0.003802164690569043, + 0.004026270937174559, + 0.0037940549664199352, + 0.003156263381242752, + 0.003151336219161749, + 0.003395374631509185, + 0.0025756626855582, + 0.0030073479283601046, + 0.003276379080489278, + 0.0037623730022460222, + 0.002838465850800276, + 0.0036448061000555754, + 0.005494415760040283, + 0.0038345814682543278, + 0.002761856187134981, + 0.004096917808055878, + 0.041439466178417206, + 0.004501369781792164, + 0.03922506794333458, + 0.01690741442143917, + 0.011391855776309967, + 0.018724385648965836, + 0.017233042046427727, + 0.01426622737199068, + 0.02378050982952118, + 0.012658420950174332, + 0.03230458125472069, + 0.009528149850666523, + 0.016292352229356766, + 0.011947660706937313, + 0.008000172674655914, + 0.008999785408377647, + 0.08627530187368393, + 0.008322720415890217, + 0.08969578891992569, + 0.0011687774676829576, + 0.002047858200967312, + 0.0020807236433029175, + 0.002465242752805352, + 0.0015178917674347758, + 0.002051733899861574, + 0.0017199038993567228, + 0.0020495858043432236, + 0.0013585839187726378, + 0.0013150946469977498, + 0.001991055905818939, + 0.0025512746069580317, + 0.003147862385958433, + 0.005476977676153183, + 0.00250004674308002, + 0.008475624956190586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.012576202861964703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059515379369259 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.02059515379369259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059515379369259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09359875321388245, + 0.0321495495736599, + 0.4296879172325134, + 0.07502997666597366, + 0.03076648898422718, + 0.01983889751136303, + 0.07359305024147034, + 0.1534341722726822, + 0.04336072504520416, + 0.04854046180844307 + ], + "credit_map": [ + 0.00716477632522583, + 0.010947315953671932, + 0.01634417288005352, + 0.014874793589115143, + 0.008035960607230663, + 0.010061082430183887, + 0.014929672703146935, + 0.011240974068641663, + 0.0025681022088974714, + 0.003342358861118555, + 0.00424780510365963, + 0.006478617433458567, + 0.002651296090334654, + 0.004140838049352169, + 0.004344499204307795, + 0.004376030992716551, + 0.01147437747567892, + 0.07067528367042542, + 0.010069193318486214, + 0.006081555970013142, + 0.009620998054742813, + 0.19164347648620605, + 0.004170202184468508, + 0.1259528398513794, + 0.007905501872301102, + 0.007962699979543686, + 0.01274476945400238, + 0.01233270950615406, + 0.007416983600705862, + 0.011105536483228207, + 0.007720352616161108, + 0.007841411046683788, + 0.004138869699090719, + 0.00379192060790956, + 0.003358788788318634, + 0.0029350058175623417, + 0.0024848359171301126, + 0.005120060406625271, + 0.004093660973012447, + 0.0048433467745780945, + 0.0017735047731548548, + 0.001812585978768766, + 0.002168676583096385, + 0.0017552226781845093, + 0.0016619113739579916, + 0.0038638717960566282, + 0.002506020711734891, + 0.004297101870179176, + 0.007644096855074167, + 0.005499720107764006, + 0.008219092153012753, + 0.0067807454615831375, + 0.007990513928234577, + 0.012949864380061626, + 0.007897067815065384, + 0.016611946746706963, + 0.005577475298196077, + 0.009526905603706837, + 0.007094441447407007, + 0.004516979213804007, + 0.005340285133570433, + 0.057085923850536346, + 0.005090431310236454, + 0.05920172855257988, + 0.003562014317139983, + 0.006190882995724678, + 0.006634156219661236, + 0.00702601857483387, + 0.0035583495628088713, + 0.007086440455168486, + 0.00335109094157815, + 0.005951773840934038, + 0.002670806599780917, + 0.0064781964756548405, + 0.004585073795169592, + 0.005670442711561918, + 0.004651757888495922, + 0.008215568959712982, + 0.0051772757433354855, + 0.011091337539255619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.02059515379369259 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741517689079046 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741517689079046, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741517689079046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14605312049388885, + 0.09973005205392838, + 0.28110983967781067, + 0.03150142729282379, + 0.044762834906578064, + 0.053294554352760315, + 0.07673671096563339, + 0.16144207119941711, + 0.06323107331991196, + 0.04213826730847359 + ], + "credit_map": [ + 0.01217232272028923, + 0.016779296100139618, + 0.021443404257297516, + 0.02356765791773796, + 0.01330715324729681, + 0.020401490852236748, + 0.018749767914414406, + 0.019632045179605484, + 0.011275473050773144, + 0.013038416393101215, + 0.012810860760509968, + 0.013682284392416477, + 0.005903373006731272, + 0.01788620464503765, + 0.0065176719799637794, + 0.018615776672959328, + 0.008147377520799637, + 0.045256033539772034, + 0.006600625347346067, + 0.004487637430429459, + 0.006875172723084688, + 0.12417533248662949, + 0.0032236415427178144, + 0.08234401047229767, + 0.0029541112016886473, + 0.0030155791901052, + 0.00684037571772933, + 0.005444821435958147, + 0.0029024137184023857, + 0.0036839989479631186, + 0.002872021635994315, + 0.0037881098687648773, + 0.007691954728215933, + 0.006192310247570276, + 0.004355583805590868, + 0.003504253691062331, + 0.004249201621860266, + 0.006068872753530741, + 0.0041069951839745045, + 0.008593665435910225, + 0.0015996526926755905, + 0.0029312625993043184, + 0.002076639560982585, + 0.0011401213705539703, + 0.0019217977533116937, + 0.021054474636912346, + 0.002216986147686839, + 0.02035362459719181, + 0.005776291713118553, + 0.004404391627758741, + 0.006133905611932278, + 0.004019331187009811, + 0.006346513982862234, + 0.023229165002703667, + 0.006049516145139933, + 0.02077759988605976, + 0.0070478832349181175, + 0.012639008462429047, + 0.008440452627837658, + 0.00640977593138814, + 0.0066652558743953705, + 0.05469094589352608, + 0.006249528378248215, + 0.05929924175143242, + 0.004809470381587744, + 0.008516594767570496, + 0.009767814539372921, + 0.01032829936593771, + 0.005570481531322002, + 0.010266795754432678, + 0.005057874601334333, + 0.008913745172321796, + 0.002661109436303377, + 0.004240185953676701, + 0.0052020736038684845, + 0.006975735072046518, + 0.004900689236819744, + 0.006508783437311649, + 0.0032277435529977083, + 0.008421948179602623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.006741517689079046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands with long fingers and neatly manicured nails visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small stack of antique leather-bound books visible, one deep burgundy and one forest green with faded gold leaf filigree on the spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands sliding the books into the last empty space on a dark, polished mahogany bookshelf crowded with similar books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm, low-key light from an off-camera source casting a soft golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin as a medium close-up angled slightly downward and execute a slow, steady zoom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897435128688812 + ], + "opposite_logprob_tokens": [ + -2.5789742469787598 + ], + "logprob_reward": -0.07897435128688812, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897435128688812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06882530450820923, + 0.01828492060303688, + 0.10552957653999329, + 0.021764159202575684, + 0.024403057992458344, + 0.16894009709358215, + 0.1465076059103012, + 0.36703598499298096, + 0.05584750697016716, + 0.022861842066049576 + ], + "credit_map": [ + 0.006662124767899513, + 0.007902671582996845, + 0.012364325113594532, + 0.013831648044288158, + 0.0056681083515286446, + 0.007353872526437044, + 0.009045002050697803, + 0.005997552536427975, + 0.001834004302509129, + 0.0020741403568536043, + 0.0034855378326028585, + 0.002668430097401142, + 0.001572754350490868, + 0.0022450622636824846, + 0.001615661894902587, + 0.002789329271763563, + 0.007796746212989092, + 0.03726385906338692, + 0.005575176328420639, + 0.007087981794029474, + 0.004713739268481731, + 0.03018787130713463, + 0.004538960754871368, + 0.008365240879356861, + 0.0020920259412378073, + 0.003132231766358018, + 0.002545183291658759, + 0.0029954644851386547, + 0.0024986881762742996, + 0.002530224621295929, + 0.0031858624424785376, + 0.0027844791766256094, + 0.0038153021596372128, + 0.0031991770956665277, + 0.0030304433312267065, + 0.0027690078131854534, + 0.0020978450775146484, + 0.002811071928590536, + 0.0035525646526366472, + 0.0031276450026780367, + 0.005352421198040247, + 0.008477807976305485, + 0.006203468423336744, + 0.004015510436147451, + 0.005922749172896147, + 0.0684649646282196, + 0.0065354108810424805, + 0.063967764377594, + 0.011121021583676338, + 0.007457032334059477, + 0.011675584129989147, + 0.009084537625312805, + 0.013129591010510921, + 0.045780397951602936, + 0.01055344007909298, + 0.037706002593040466, + 0.01566733606159687, + 0.027139125391840935, + 0.020678777247667313, + 0.015018252655863762, + 0.013887738808989525, + 0.12646494805812836, + 0.013033609837293625, + 0.13514617085456848, + 0.004872458986938, + 0.008070273324847221, + 0.007645349018275738, + 0.007962621748447418, + 0.004023063462227583, + 0.010718300007283688, + 0.0035961088724434376, + 0.008959331549704075, + 0.0018493118695914745, + 0.0036843554116785526, + 0.0035334364511072636, + 0.0037563482765108347, + 0.0019187707221135497, + 0.003117998130619526, + 0.002344971289858222, + 0.002656647702679038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07897435128688812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_tensor([4517, 4517])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_tensor([4517, 4517])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85e9e6137a2113b1488509858877596ee9c1c0ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_tensor([4517, 4517])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An uninterrupted single shot, filmed with the quality of a nature documentary, slowly advances from a low angle, deep into a damp woodland. A massive monitor lizard crouches amidst rotting leaves and dark green moss. Its body, covered in ancient, beaded scales, is a dull earthy gray, its rough skin, like weathered armor, its uneven texture revealed in the dappled light filtering through the dense canopy. Its powerful limbs are firmly rooted in the moist earth, its body motionless, like a prehistoric sculpture. The only movement comes from its long, forked yellow tongue, which rapidly extends and retracts, keenly gathering information from the earthy air. As the camera draws closer, almost touching its cold, impassive black eyes, the stark contrast between stillness and movement heightens the primal, poised tension of the hunt, filling the scene with a breathtaking sense of anticipation.", + "seed": 4809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_4517_seed_4809_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03006463828496635, + "vlm_logprob": -0.03006463828496635, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03006463828496635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024195391684770584 + ], + "opposite_logprob_tokens": [ + -3.899195432662964 + ], + "logprob_reward": -0.024195391684770584, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024195391684770584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23162545263767242, + 0.10216289758682251, + 0.1681116223335266, + 0.15898190438747406, + 0.048470959067344666, + 0.031081654131412506, + 0.05200360342860222, + 0.07486472278833389, + 0.02636886015534401, + 0.1063283309340477 + ], + "credit_map": [ + 0.023907382041215897, + 0.029612159356474876, + 0.04742986708879471, + 0.026735743507742882, + 0.027040282264351845, + 0.0207792017608881, + 0.02868512086570263, + 0.027435705065727234, + 0.013520127162337303, + 0.007636663969606161, + 0.031834036111831665, + 0.011051937006413937, + 0.008775054477155209, + 0.009080946445465088, + 0.009794725105166435, + 0.010469415225088596, + 0.032359689474105835, + 0.010254145599901676, + 0.013321091420948505, + 0.025025593116879463, + 0.01813744381070137, + 0.04162457585334778, + 0.014451065100729465, + 0.012938017025589943, + 0.029615560546517372, + 0.010485846549272537, + 0.010594903491437435, + 0.02681124024093151, + 0.03769151121377945, + 0.013522772118449211, + 0.018184874206781387, + 0.012075192295014858, + 0.007219566032290459, + 0.005049581173807383, + 0.004977491218596697, + 0.006057385820895433, + 0.011231803335249424, + 0.004799002781510353, + 0.0053789992816746235, + 0.003757128259167075, + 0.0046530128456652164, + 0.004072338342666626, + 0.0034776669926941395, + 0.00429295189678669, + 0.002788012148812413, + 0.0031794526148587465, + 0.003301348304376006, + 0.005316872149705887, + 0.011489476077258587, + 0.004034288693219423, + 0.0031272645574063063, + 0.01616460457444191, + 0.004441636614501476, + 0.003334402572363615, + 0.005434277001768351, + 0.003977654036134481, + 0.013030961155891418, + 0.004547736141830683, + 0.0033974747639149427, + 0.00872763805091381, + 0.007294371258467436, + 0.023282404989004135, + 0.007751693483442068, + 0.006832443177700043, + 0.003859084565192461, + 0.002763505559414625, + 0.002935636555776, + 0.0035442793741822243, + 0.002646252978593111, + 0.004638919606804848, + 0.0028070833068341017, + 0.0031741010025143623, + 0.02706151455640793, + 0.007691224571317434, + 0.0074636065401136875, + 0.010283317416906357, + 0.00794603955000639, + 0.008610664866864681, + 0.011310484260320663, + 0.02596147358417511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.024195391684770584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032025498803704977 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.0032025498803704977, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032025498803704977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21063236892223358, + 0.08449042588472366, + 0.15432602167129517, + 0.11546193063259125, + 0.08604265749454498, + 0.10817153006792068, + 0.027842050418257713, + 0.03585534915328026, + 0.07346132397651672, + 0.10371634364128113 + ], + "credit_map": [ + 0.025875316932797432, + 0.012622939422726631, + 0.07194147258996964, + 0.01648944430053234, + 0.008118974976241589, + 0.011013845913112164, + 0.018859481438994408, + 0.04571089521050453, + 0.0171718280762434, + 0.005587581545114517, + 0.028484243899583817, + 0.00827147625386715, + 0.006247951183468103, + 0.005333141423761845, + 0.007655678316950798, + 0.005738522857427597, + 0.0406380221247673, + 0.006405690219253302, + 0.00949756521731615, + 0.013585105538368225, + 0.010007843375205994, + 0.04983770102262497, + 0.01211306732147932, + 0.01224101148545742, + 0.02363603748381138, + 0.013634766452014446, + 0.012899542227387428, + 0.0176315288990736, + 0.016746046021580696, + 0.009743615053594112, + 0.01213392149657011, + 0.0090364720672369, + 0.014323881827294827, + 0.005923941731452942, + 0.006377221085131168, + 0.008373800665140152, + 0.026392830535769463, + 0.006989848334342241, + 0.007701178081333637, + 0.009959954768419266, + 0.011353053152561188, + 0.008599274791777134, + 0.00672810198739171, + 0.02421436831355095, + 0.027557460591197014, + 0.006875285413116217, + 0.010836205445230007, + 0.012007786892354488, + 0.0033025764860212803, + 0.005123026669025421, + 0.002378372475504875, + 0.004250444006174803, + 0.002765413373708725, + 0.0024322078097611666, + 0.0032885742839425802, + 0.004301436711102724, + 0.004351044539362192, + 0.004017445258796215, + 0.0031598592177033424, + 0.006410653702914715, + 0.0029854937456548214, + 0.007680725771933794, + 0.003751946147531271, + 0.0034981798380613327, + 0.012965293601155281, + 0.005142240785062313, + 0.004585243295878172, + 0.0073680579662323, + 0.010112021118402481, + 0.019691316410899162, + 0.007123631425201893, + 0.006473518908023834, + 0.029879016801714897, + 0.00927711185067892, + 0.00793954636901617, + 0.009424564428627491, + 0.006051126401871443, + 0.006112587172538042, + 0.009342976845800877, + 0.025689411908388138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0032025498803704977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962086841464043 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005962086841464043, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962086841464043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19883476197719574, + 0.1264118105173111, + 0.07578849792480469, + 0.0541745126247406, + 0.03499438986182213, + 0.06603462994098663, + 0.08721479773521423, + 0.09315083920955658, + 0.2032277137041092, + 0.06016800180077553 + ], + "credit_map": [ + 0.022720828652381897, + 0.019022222608327866, + 0.057963043451309204, + 0.019475150853395462, + 0.020057568326592445, + 0.01187928020954132, + 0.023649906739592552, + 0.024066757410764694, + 0.011772241443395615, + 0.009982757270336151, + 0.018695969134569168, + 0.015993373468518257, + 0.01045871153473854, + 0.015156548470258713, + 0.014889094047248363, + 0.029463106766343117, + 0.014398735947906971, + 0.010431939736008644, + 0.008354132995009422, + 0.008998783305287361, + 0.00868432130664587, + 0.006472268141806126, + 0.009939619340002537, + 0.008508697152137756, + 0.006542775314301252, + 0.005832153372466564, + 0.006335902959108353, + 0.01186333131045103, + 0.005244103260338306, + 0.0070310840383172035, + 0.006212317384779453, + 0.005112848244607449, + 0.005003047175705433, + 0.004272181540727615, + 0.0038418984040617943, + 0.005946335848420858, + 0.004842681810259819, + 0.003881208598613739, + 0.0036945315077900887, + 0.0035125052090734243, + 0.007970486767590046, + 0.00447836285457015, + 0.004045424982905388, + 0.01467701606452465, + 0.018473075702786446, + 0.004213342908769846, + 0.006165234372019768, + 0.006011693738400936, + 0.005486627575010061, + 0.0070490711368620396, + 0.003808234818279743, + 0.03203965350985527, + 0.019107308238744736, + 0.004379917401820421, + 0.008870262652635574, + 0.006473717279732227, + 0.010292427614331245, + 0.008715922944247723, + 0.0050246804021298885, + 0.020067738369107246, + 0.005826666951179504, + 0.027655795216560364, + 0.007281470112502575, + 0.008286143653094769, + 0.03416365757584572, + 0.016100574284791946, + 0.014081425033509731, + 0.021179920062422752, + 0.015272913500666618, + 0.06820657104253769, + 0.01731177791953087, + 0.01691087707877159, + 0.006151549983769655, + 0.007088434882462025, + 0.00979849137365818, + 0.006443124730139971, + 0.007600707467645407, + 0.006656476762145758, + 0.009059974923729897, + 0.007369241211563349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005962086841464043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941095411777496 + ], + "opposite_logprob_tokens": [ + -2.929941177368164 + ], + "logprob_reward": -0.054941095411777496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941095411777496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09641905128955841, + 0.34804028272628784, + 0.19941192865371704, + 0.05928520858287811, + 0.04948430508375168, + 0.06140214949846268, + 0.024873003363609314, + 0.049099575728178024, + 0.03977218270301819, + 0.07221224904060364 + ], + "credit_map": [ + 0.018441636115312576, + 0.012616061605513096, + 0.014323251321911812, + 0.010676775127649307, + 0.008092815987765789, + 0.007625378668308258, + 0.011591152288019657, + 0.013051988556981087, + 0.04693396016955376, + 0.030509209260344505, + 0.09732189774513245, + 0.0321391299366951, + 0.04096735268831253, + 0.030928583815693855, + 0.039685510098934174, + 0.029554663226008415, + 0.051530953496694565, + 0.01028689555823803, + 0.011934450827538967, + 0.026127053424715996, + 0.019401010125875473, + 0.05169817432761192, + 0.016175799071788788, + 0.012257580645382404, + 0.01252049021422863, + 0.0051275985315442085, + 0.005525499582290649, + 0.009495471604168415, + 0.009400769136846066, + 0.006424570456147194, + 0.006210209336131811, + 0.00458059785887599, + 0.00841766968369484, + 0.00416272459551692, + 0.004243984818458557, + 0.006335260346531868, + 0.01234593614935875, + 0.005087890196591616, + 0.004137319978326559, + 0.004753517918288708, + 0.012207095511257648, + 0.006709353998303413, + 0.005379832349717617, + 0.008126500993967056, + 0.01311584748327732, + 0.003355776658281684, + 0.005889034830033779, + 0.006618707440793514, + 0.0026755451690405607, + 0.002864382928237319, + 0.002185804070904851, + 0.0037733011413365602, + 0.0031944436486810446, + 0.002591565018519759, + 0.0032580713741481304, + 0.0043298895470798016, + 0.008318392559885979, + 0.0062631904147565365, + 0.0038870777934789658, + 0.007204817607998848, + 0.005310529377311468, + 0.008934053592383862, + 0.004512320272624493, + 0.004669192247092724, + 0.006754618138074875, + 0.00490550184622407, + 0.004477232228964567, + 0.005476274061948061, + 0.0047922092489898205, + 0.004645250272005796, + 0.005053716246038675, + 0.003667385783046484, + 0.015149435959756374, + 0.006513210013508797, + 0.009053446352481842, + 0.006185485515743494, + 0.007070042658597231, + 0.004716916475445032, + 0.008678747341036797, + 0.014844962395727634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.054941095411777496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202206760644913 + ], + "opposite_logprob_tokens": [ + -2.8120219707489014 + ], + "logprob_reward": -0.06202206760644913, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202206760644913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13712601363658905, + 0.04050200432538986, + 0.12275011092424393, + 0.045406848192214966, + 0.057997945696115494, + 0.06672646850347519, + 0.08815769106149673, + 0.09916473180055618, + 0.15328766405582428, + 0.1888805329799652 + ], + "credit_map": [ + 0.01390024833381176, + 0.012205857783555984, + 0.03874418884515762, + 0.012286313809454441, + 0.01174249965697527, + 0.00800828356295824, + 0.016584208235144615, + 0.023654408752918243, + 0.00394066795706749, + 0.0036199286114424467, + 0.004260791931301355, + 0.00655136676505208, + 0.005520198959857225, + 0.005204466171562672, + 0.005525821354240179, + 0.005878761410713196, + 0.03540171682834625, + 0.00671191094443202, + 0.004908205009996891, + 0.011767568066716194, + 0.007688742130994797, + 0.038111597299575806, + 0.009178399108350277, + 0.00898196455091238, + 0.01085594017058611, + 0.0033268281258642673, + 0.0031669698655605316, + 0.007194732315838337, + 0.008883800357580185, + 0.0032022823579609394, + 0.004721010569483042, + 0.004055289085954428, + 0.005990868899971247, + 0.005193240940570831, + 0.005597705952823162, + 0.00876829493790865, + 0.014966130256652832, + 0.0051958016119897366, + 0.00672049168497324, + 0.005565414670854807, + 0.005148714873939753, + 0.006376159843057394, + 0.005751246586441994, + 0.018178347498178482, + 0.012365374714136124, + 0.003789372742176056, + 0.007579685188829899, + 0.0075375670567154884, + 0.007151729427278042, + 0.005947940517216921, + 0.0049440450966358185, + 0.03282969072461128, + 0.01667112670838833, + 0.004719447810202837, + 0.009620885364711285, + 0.006272828206419945, + 0.014950963668525219, + 0.008911648765206337, + 0.004892793484032154, + 0.01568780466914177, + 0.006559181958436966, + 0.03283029794692993, + 0.008444393053650856, + 0.00688764825463295, + 0.021002624183893204, + 0.01557273231446743, + 0.013539749197661877, + 0.015711447224020958, + 0.013448411598801613, + 0.048960134387016296, + 0.012219934724271297, + 0.012832625769078732, + 0.04923374578356743, + 0.014941589906811714, + 0.016273552551865578, + 0.018143784254789352, + 0.013873576186597347, + 0.016440987586975098, + 0.021198377013206482, + 0.03877493366599083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.06202206760644913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted single shot, filmed with the quality of a nature documentary, slowly advances from a low angle, deep into a damp woodland. A massive monitor lizard crouches amidst rotting leaves and dark green moss. Its body, covered in ancient, beaded scales, is a dull earthy gray, its rough skin, like weathered armor, its uneven texture revealed in the dappled light filtering through the dense canopy. Its powerful limbs are firmly rooted in the moist earth, its body motionless, like a prehistoric sculpture. The only movement comes from its long, forked yellow tongue, which rapidly extends and retracts, keenly gathering information from the earthy air. As the camera draws closer, almost touching its cold, impassive black eyes, the stark contrast between stillness and movement heightens the primal, poised tension of the hunt, filling the scene with a breathtaking sense of anticipation.", + "seed": 4810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_4517_seed_4810_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05210292087867856, + "vlm_logprob": -0.05210292087867856, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05210292087867856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01574055291712284 + ], + "opposite_logprob_tokens": [ + -4.390740394592285 + ], + "logprob_reward": -0.01574055291712284, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01574055291712284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36356303095817566, + 0.12815959751605988, + 0.049858007580041885, + 0.05151429772377014, + 0.12613371014595032, + 0.14661318063735962, + 0.03911169618368149, + 0.03311462327837944, + 0.03255261853337288, + 0.02937924489378929 + ], + "credit_map": [ + 0.037250012159347534, + 0.029033292084932327, + 0.07297185063362122, + 0.03159336745738983, + 0.03498284891247749, + 0.039936363697052, + 0.0907476395368576, + 0.027047643437981606, + 0.01658691093325615, + 0.0143691161647439, + 0.013829939998686314, + 0.017012707889080048, + 0.015554726123809814, + 0.0197538360953331, + 0.01947668194770813, + 0.011575683951377869, + 0.004131604917347431, + 0.00529186986386776, + 0.0052183689549565315, + 0.005436871666461229, + 0.006735167931765318, + 0.008602006360888481, + 0.00764573086053133, + 0.006796382367610931, + 0.004191802814602852, + 0.005674160085618496, + 0.005529440939426422, + 0.006871235091239214, + 0.007861826568841934, + 0.008566711097955704, + 0.007841464132070541, + 0.004977659787982702, + 0.010203184559941292, + 0.009449099190533161, + 0.02634619176387787, + 0.011740922927856445, + 0.013993402011692524, + 0.015447065234184265, + 0.023451948538422585, + 0.015501904301345348, + 0.012505270540714264, + 0.0071840048767626286, + 0.05394772067666054, + 0.00887124054133892, + 0.006432500667870045, + 0.00874731969088316, + 0.04034590721130371, + 0.00857921689748764, + 0.004094153642654419, + 0.0045793005265295506, + 0.005540815647691488, + 0.005102416500449181, + 0.003328808583319187, + 0.006079548969864845, + 0.005337978713214397, + 0.005048672202974558, + 0.003965675365179777, + 0.004458112642168999, + 0.003096533240750432, + 0.004676292650401592, + 0.005497799254953861, + 0.003970127087086439, + 0.0044122436083853245, + 0.003037839662283659, + 0.004006541334092617, + 0.0040927063673734665, + 0.005334043875336647, + 0.004197596572339535, + 0.0044217826798558235, + 0.0038431785069406033, + 0.0039746444672346115, + 0.00268212310038507, + 0.0038588917814195156, + 0.003604640020057559, + 0.003540133126080036, + 0.004680372308939695, + 0.004653688054531813, + 0.002701910911127925, + 0.0037031807005405426, + 0.002636425429955125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01574055291712284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003200292121618986 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003200292121618986, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003200292121618986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4066700339317322, + 0.11829265207052231, + 0.029796961694955826, + 0.029252754524350166, + 0.04811292141675949, + 0.17011624574661255, + 0.0574774295091629, + 0.041291240602731705, + 0.050770312547683716, + 0.048219479620456696 + ], + "credit_map": [ + 0.03521955385804176, + 0.025344738736748695, + 0.09749377518892288, + 0.035599369555711746, + 0.028290649875998497, + 0.03648821637034416, + 0.11030109971761703, + 0.037932634353637695, + 0.017356887459754944, + 0.007519616279751062, + 0.018718207255005836, + 0.013200045563280582, + 0.009735321626067162, + 0.016658613458275795, + 0.018156001344323158, + 0.01694795861840248, + 0.0035681347362697124, + 0.002590959193184972, + 0.0035309188533574343, + 0.00320641603320837, + 0.0046453275717794895, + 0.003382869064807892, + 0.004280848428606987, + 0.004591486882418394, + 0.004056329373270273, + 0.003644414944574237, + 0.004058602266013622, + 0.0033520180732011795, + 0.0038772791158407927, + 0.0036071669310331345, + 0.0038994455244392157, + 0.002757497364655137, + 0.004781167954206467, + 0.0028144042007625103, + 0.01289303507655859, + 0.005536351352930069, + 0.004329894203692675, + 0.004124731291085482, + 0.008037702180445194, + 0.005595633294433355, + 0.01105274073779583, + 0.005223578307777643, + 0.04897240921854973, + 0.00946769118309021, + 0.008769266307353973, + 0.018998220562934875, + 0.05950946733355522, + 0.008122887462377548, + 0.007628781255334616, + 0.005762161687016487, + 0.008608200587332249, + 0.0070061939768493176, + 0.0036668728571385145, + 0.009370803833007812, + 0.009084617719054222, + 0.006349795963615179, + 0.004804777912795544, + 0.004423779435455799, + 0.0059901271015405655, + 0.004794910084456205, + 0.004122656304389238, + 0.005496194586157799, + 0.0068718502297997475, + 0.004786946345120668, + 0.009682651609182358, + 0.005048766266554594, + 0.007300945930182934, + 0.004850299563258886, + 0.004244998563081026, + 0.00870441272854805, + 0.006672730669379234, + 0.004265506286174059, + 0.007766112219542265, + 0.004157704766839743, + 0.00793550070375204, + 0.0047382391057908535, + 0.003981647547334433, + 0.006902018561959267, + 0.007028417196124792, + 0.005709839519113302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.003200292121618986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979746226221323 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.00979746226221323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979746226221323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25355902314186096, + 0.11095434427261353, + 0.055905211716890335, + 0.04099112004041672, + 0.10971321165561676, + 0.10915839672088623, + 0.07539056241512299, + 0.08935699611902237, + 0.06844928115606308, + 0.08652182668447495 + ], + "credit_map": [ + 0.03669179603457451, + 0.016940589994192123, + 0.05167264863848686, + 0.028727324679493904, + 0.01921921595931053, + 0.029235245659947395, + 0.04942164942622185, + 0.021650530397892, + 0.014075773768126965, + 0.011640936136245728, + 0.018425872549414635, + 0.012257277965545654, + 0.01153153833001852, + 0.01218643132597208, + 0.015884922817349434, + 0.014951578341424465, + 0.0059478385373950005, + 0.005135284271091223, + 0.006437891628593206, + 0.0060407803393900394, + 0.008421573787927628, + 0.008987197652459145, + 0.007950936444103718, + 0.006983703467994928, + 0.004605281166732311, + 0.004082502331584692, + 0.00437348335981369, + 0.003489943454042077, + 0.005445162300020456, + 0.007867580279707909, + 0.005977352149784565, + 0.005149807780981064, + 0.009861359372735023, + 0.006946384906768799, + 0.010468095541000366, + 0.008029495365917683, + 0.007921145297586918, + 0.02376031130552292, + 0.03428974002599716, + 0.008436664938926697, + 0.007242597173899412, + 0.004835333209484816, + 0.022219378501176834, + 0.008132146671414375, + 0.007708316203206778, + 0.01440400443971157, + 0.03645537048578262, + 0.008161249570548534, + 0.007324352860450745, + 0.0071462602354586124, + 0.009157287888228893, + 0.011294971220195293, + 0.0064187804237008095, + 0.011320778168737888, + 0.011210955679416656, + 0.0115171754732728, + 0.008866998367011547, + 0.009817116893827915, + 0.01080078724771738, + 0.0120421526953578, + 0.011808237060904503, + 0.01083612721413374, + 0.013226122595369816, + 0.011959448456764221, + 0.010337728075683117, + 0.007598537486046553, + 0.009419210255146027, + 0.006598061416298151, + 0.008531575091183186, + 0.009078726172447205, + 0.008888286538422108, + 0.007997152395546436, + 0.010552680119872093, + 0.011407075449824333, + 0.009064218029379845, + 0.011799314990639687, + 0.014255116693675518, + 0.009559796191751957, + 0.011416524648666382, + 0.008467093110084534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00979746226221323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900478482246399 + ], + "opposite_logprob_tokens": [ + -2.4640047550201416 + ], + "logprob_reward": -0.08900478482246399, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900478482246399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5035444498062134, + 0.1072421744465828, + 0.06342815607786179, + 0.038938939571380615, + 0.12041191011667252, + 0.04833979532122612, + 0.04739747196435928, + 0.0231575146317482, + 0.022773174569010735, + 0.024766376242041588 + ], + "credit_map": [ + 0.042044855654239655, + 0.04045889154076576, + 0.09788742661476135, + 0.03894883394241333, + 0.05060212314128876, + 0.05092444270849228, + 0.12866906821727753, + 0.054008789360523224, + 0.011347210966050625, + 0.013937151990830898, + 0.009388030506670475, + 0.012781139463186264, + 0.019864307716488838, + 0.011906556785106659, + 0.016557853668928146, + 0.011459925211966038, + 0.006289966404438019, + 0.008182177320122719, + 0.005411809775978327, + 0.0071748350746929646, + 0.010548148304224014, + 0.009229484014213085, + 0.010280114598572254, + 0.006311617325991392, + 0.0038886163383722305, + 0.00456001004204154, + 0.0033397965598851442, + 0.0034415361005812883, + 0.006871209014207125, + 0.005799471400678158, + 0.006007814779877663, + 0.005030485335737467, + 0.008753366768360138, + 0.012115227058529854, + 0.030008573085069656, + 0.008686288259923458, + 0.008777763694524765, + 0.010625568218529224, + 0.034226931631565094, + 0.007218188140541315, + 0.004636144731193781, + 0.002979598706588149, + 0.012037092819809914, + 0.0025307913310825825, + 0.003685861360281706, + 0.005262361373752356, + 0.01377463061362505, + 0.003433310892432928, + 0.003935924265533686, + 0.0038935495540499687, + 0.010740292258560658, + 0.004754678346216679, + 0.005645918659865856, + 0.004947526380419731, + 0.00717571284621954, + 0.006303868256509304, + 0.0030119072180241346, + 0.0022244900465011597, + 0.0024674157612025738, + 0.003936904016882181, + 0.00225912407040596, + 0.0025217400398105383, + 0.003003346733748913, + 0.00373258744366467, + 0.002592474455013871, + 0.0024235283490270376, + 0.003739919513463974, + 0.003019060241058469, + 0.0028884042985737324, + 0.0025299531407654285, + 0.0031399019062519073, + 0.0024399342946708202, + 0.0030902023427188396, + 0.0026516623329371214, + 0.004255347419530153, + 0.003204986685886979, + 0.002034286269918084, + 0.002971322275698185, + 0.003594170091673732, + 0.0029643988236784935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.08900478482246399 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277151226997375 + ], + "opposite_logprob_tokens": [ + -2.0177714824676514 + ], + "logprob_reward": -0.14277151226997375, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277151226997375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3167099356651306, + 0.11955858021974564, + 0.04853975027799606, + 0.039853956550359726, + 0.09445016831159592, + 0.13256070017814636, + 0.07086268812417984, + 0.05989852175116539, + 0.029204826802015305, + 0.08836085349321365 + ], + "credit_map": [ + 0.042581673711538315, + 0.027165008708834648, + 0.058256253600120544, + 0.03526528924703598, + 0.03076413832604885, + 0.029325922951102257, + 0.06526097655296326, + 0.028090650215744972, + 0.013858391903340816, + 0.01157338172197342, + 0.023488840088248253, + 0.015347396954894066, + 0.00918280053883791, + 0.014939013868570328, + 0.01768375188112259, + 0.013485007919371128, + 0.005647710524499416, + 0.004728652071207762, + 0.0050954814068973064, + 0.00628211023285985, + 0.007862698286771774, + 0.007616130635142326, + 0.006694646552205086, + 0.004612320102751255, + 0.004933272488415241, + 0.003427480347454548, + 0.003522810759022832, + 0.004295813851058483, + 0.005915715359151363, + 0.007866261526942253, + 0.005908177699893713, + 0.003984426148235798, + 0.007917887531220913, + 0.005918986164033413, + 0.013490583747625351, + 0.005896895658224821, + 0.006771904416382313, + 0.017494799569249153, + 0.0280742347240448, + 0.008884877897799015, + 0.008012800477445126, + 0.0037405493203550577, + 0.03847985342144966, + 0.008921285159885883, + 0.005706732161343098, + 0.0143344821408391, + 0.046426527202129364, + 0.006938467733561993, + 0.006552353035658598, + 0.006359168794006109, + 0.008883790113031864, + 0.009073585271835327, + 0.006720677483826876, + 0.010079552419483662, + 0.012405182234942913, + 0.01078837737441063, + 0.006880729924887419, + 0.006549235433340073, + 0.00826443824917078, + 0.007737722247838974, + 0.007206386886537075, + 0.007938999682664871, + 0.008272500708699226, + 0.00704850722104311, + 0.00465357955545187, + 0.0032561139669269323, + 0.00367360794916749, + 0.003035747678950429, + 0.0028085808735340834, + 0.0037939981557428837, + 0.0038529904559254646, + 0.004130207933485508, + 0.013326553627848625, + 0.00913842674344778, + 0.009927570819854736, + 0.010932297445833683, + 0.012567036785185337, + 0.013098745606839657, + 0.010958751663565636, + 0.008411468006670475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.14277151226997375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted single shot, filmed with the quality of a nature documentary, slowly advances from a low angle, deep into a damp woodland. A massive monitor lizard crouches amidst rotting leaves and dark green moss. Its body, covered in ancient, beaded scales, is a dull earthy gray, its rough skin, like weathered armor, its uneven texture revealed in the dappled light filtering through the dense canopy. Its powerful limbs are firmly rooted in the moist earth, its body motionless, like a prehistoric sculpture. The only movement comes from its long, forked yellow tongue, which rapidly extends and retracts, keenly gathering information from the earthy air. As the camera draws closer, almost touching its cold, impassive black eyes, the stark contrast between stillness and movement heightens the primal, poised tension of the hunt, filling the scene with a breathtaking sense of anticipation.", + "seed": 4811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_4517_seed_4811_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03565481370314956, + "vlm_logprob": -0.03565481370314956, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03565481370314956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1021856963634491 + ], + "opposite_logprob_tokens": [ + -2.3521857261657715 + ], + "logprob_reward": -0.1021856963634491, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1021856963634491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08360160887241364, + 0.18934307992458344, + 0.047440048307180405, + 0.03335755690932274, + 0.16451582312583923, + 0.11062859743833542, + 0.03292570263147354, + 0.16571763157844543, + 0.0823441669344902, + 0.09012576192617416 + ], + "credit_map": [ + 0.009461579844355583, + 0.012179347686469555, + 0.010341186076402664, + 0.00970954354852438, + 0.01200496032834053, + 0.006610719952732325, + 0.013456803746521473, + 0.009837468154728413, + 0.01842339150607586, + 0.032979752868413925, + 0.01920192874968052, + 0.03506189212203026, + 0.014494763687252998, + 0.02326931618154049, + 0.01975059323012829, + 0.0261614378541708, + 0.006125978659838438, + 0.007285318337380886, + 0.009602737613022327, + 0.0042047640308737755, + 0.004520875867456198, + 0.0039350357837975025, + 0.004713220521807671, + 0.007052119821310043, + 0.004234124906361103, + 0.004970828536897898, + 0.003427579067647457, + 0.002507143886759877, + 0.0056676110252738, + 0.005079740192741156, + 0.003382013412192464, + 0.004088513553142548, + 0.03570560738444328, + 0.016042931005358696, + 0.017858298495411873, + 0.017444318160414696, + 0.033670574426651, + 0.011050842702388763, + 0.01515323668718338, + 0.017590010538697243, + 0.011386333964765072, + 0.006129428278654814, + 0.0076085468754172325, + 0.007411119993776083, + 0.020777760073542595, + 0.015774529427289963, + 0.028969408944249153, + 0.01257147267460823, + 0.005338744260370731, + 0.003631552215665579, + 0.00553292129188776, + 0.0035472577437758446, + 0.0032299344893544912, + 0.0037347152829170227, + 0.003419154090806842, + 0.004491421859711409, + 0.01725887879729271, + 0.018336724489927292, + 0.006679865065962076, + 0.030799755826592445, + 0.058773186057806015, + 0.015838509425520897, + 0.011048838496208191, + 0.006981878541409969, + 0.006545986980199814, + 0.011713786981999874, + 0.0076538920402526855, + 0.014410151168704033, + 0.015606936067342758, + 0.006848534569144249, + 0.011108490638434887, + 0.008456388488411903, + 0.010139397345483303, + 0.013956397771835327, + 0.009887284599244595, + 0.014313634485006332, + 0.01196854654699564, + 0.008120059967041016, + 0.012877285480499268, + 0.008863152004778385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.1021856963634491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645863547921181 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645863547921181, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645863547921181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22815477848052979, + 0.14602845907211304, + 0.09695098549127579, + 0.05851069092750549, + 0.06580543518066406, + 0.14607608318328857, + 0.054260801523923874, + 0.08273962885141373, + 0.05411225184798241, + 0.06736084073781967 + ], + "credit_map": [ + 0.02288295328617096, + 0.03030720353126526, + 0.03021235577762127, + 0.0396454855799675, + 0.03209482133388519, + 0.02169916220009327, + 0.02888106368482113, + 0.02243172749876976, + 0.017408961430191994, + 0.023134881630539894, + 0.01396961323916912, + 0.027851298451423645, + 0.014911922626197338, + 0.01484899502247572, + 0.01517424825578928, + 0.018728546798229218, + 0.010157926008105278, + 0.005278137046843767, + 0.03103587217628956, + 0.009217658080160618, + 0.005791067611426115, + 0.00675668939948082, + 0.00804727803915739, + 0.020666353404521942, + 0.007707596756517887, + 0.009219528175890446, + 0.007287606131285429, + 0.005898603703826666, + 0.009120207279920578, + 0.007177016232162714, + 0.0060663484036922455, + 0.00603377865627408, + 0.01470024324953556, + 0.006535176187753677, + 0.007442292757332325, + 0.008914613164961338, + 0.005641337018460035, + 0.006870704237371683, + 0.00870782695710659, + 0.00699323695152998, + 0.013883935287594795, + 0.00962926261126995, + 0.01127274427562952, + 0.011483744718134403, + 0.030446460470557213, + 0.02071753889322281, + 0.032769445329904556, + 0.015872962772846222, + 0.00910666212439537, + 0.007460959255695343, + 0.006739629898220301, + 0.006209033075720072, + 0.004714257083833218, + 0.006624448113143444, + 0.005625448189675808, + 0.007780363317579031, + 0.01038418896496296, + 0.01025508251041174, + 0.007720360066741705, + 0.0116195073351264, + 0.009382402524352074, + 0.013482937589287758, + 0.008652557618916035, + 0.01124258991330862, + 0.006047022994607687, + 0.005617387592792511, + 0.005218596663326025, + 0.008519339375197887, + 0.006117984652519226, + 0.007089899852871895, + 0.008658118546009064, + 0.006843901239335537, + 0.007162932772189379, + 0.00693924305960536, + 0.0074458313174545765, + 0.01773405261337757, + 0.007511443924158812, + 0.005799341946840286, + 0.008819054812192917, + 0.005948936566710472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004645863547921181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006749450694769621 + ], + "opposite_logprob_tokens": [ + -5.006749629974365 + ], + "logprob_reward": -0.006749450694769621, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749450694769621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07270835340023041, + 0.08011186122894287, + 0.08898083865642548, + 0.07252038270235062, + 0.112195685505867, + 0.0571879968047142, + 0.029279088601469994, + 0.36284151673316956, + 0.09342792630195618, + 0.0307463426142931 + ], + "credit_map": [ + 0.007822909392416477, + 0.01219172216951847, + 0.007255536038428545, + 0.00915882084518671, + 0.009365759789943695, + 0.0062810382805764675, + 0.012893933802843094, + 0.007738626562058926, + 0.00863927137106657, + 0.01382916048169136, + 0.009068384766578674, + 0.009058681316673756, + 0.00890320260077715, + 0.009303022176027298, + 0.007913258858025074, + 0.013396869413554668, + 0.009524422697722912, + 0.009818291291594505, + 0.027998289093375206, + 0.007788642775267363, + 0.005732645746320486, + 0.0053020440973341465, + 0.007427755277603865, + 0.015388733707368374, + 0.008749856613576412, + 0.022897951304912567, + 0.010173826478421688, + 0.006254792213439941, + 0.008212916553020477, + 0.007158887572586536, + 0.004485623445361853, + 0.004586527124047279, + 0.016451653093099594, + 0.018041497096419334, + 0.010610991157591343, + 0.006608036812394857, + 0.034771762788295746, + 0.008922066539525986, + 0.007317788898944855, + 0.009471881203353405, + 0.004844251088798046, + 0.0036226389929652214, + 0.0036302930675446987, + 0.007234394084662199, + 0.014866748824715614, + 0.004742370918393135, + 0.013660204596817493, + 0.004587090108543634, + 0.003858238225802779, + 0.005786656402051449, + 0.0030699684284627438, + 0.003555899951606989, + 0.0036749483551830053, + 0.0034376156982034445, + 0.003984237089753151, + 0.0019115209579467773, + 0.03626656532287598, + 0.04001054912805557, + 0.014172971248626709, + 0.07050368934869766, + 0.1288396567106247, + 0.0337674580514431, + 0.025276988744735718, + 0.014003616757690907, + 0.008745177648961544, + 0.013187631964683533, + 0.010655868798494339, + 0.012386809103190899, + 0.011690463870763779, + 0.008906198665499687, + 0.012300631031394005, + 0.015555139631032944, + 0.003512656083330512, + 0.003831782378256321, + 0.004339573439210653, + 0.004286209121346474, + 0.0030594312120229006, + 0.003126665949821472, + 0.005462936591356993, + 0.0031270834151655436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.006749450694769621 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048635248094797134 + ], + "opposite_logprob_tokens": [ + -3.048635244369507 + ], + "logprob_reward": -0.048635248094797134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048635248094797134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053879398852586746, + 0.20707780122756958, + 0.07064451277256012, + 0.05387429520487785, + 0.07638319581747055, + 0.3823220431804657, + 0.031090429052710533, + 0.05107443407177925, + 0.034905921667814255, + 0.03874797374010086 + ], + "credit_map": [ + 0.005538037978112698, + 0.0058078765869140625, + 0.004589158110320568, + 0.01210057083517313, + 0.007713815663009882, + 0.00551812257617712, + 0.006414318922907114, + 0.006197505164891481, + 0.021296506747603416, + 0.03798165172338486, + 0.01860484667122364, + 0.03810146078467369, + 0.017508652061223984, + 0.024538496509194374, + 0.02167550101876259, + 0.027370691299438477, + 0.0072145708836615086, + 0.009028254076838493, + 0.020143479108810425, + 0.007876348681747913, + 0.005029184743762016, + 0.004904440604150295, + 0.005397817585617304, + 0.011050418019294739, + 0.008611922152340412, + 0.012919524684548378, + 0.0051078787073493, + 0.003974455408751965, + 0.010162481106817722, + 0.004584935959428549, + 0.0045410809107124805, + 0.00397202093154192, + 0.010304756462574005, + 0.015529160387814045, + 0.009260368533432484, + 0.008647371083498001, + 0.008862926624715328, + 0.006918860599398613, + 0.008848385885357857, + 0.008011369965970516, + 0.04571477696299553, + 0.013757848180830479, + 0.013819582760334015, + 0.04455236345529556, + 0.0643070787191391, + 0.014344857074320316, + 0.1514027863740921, + 0.03442277014255524, + 0.003355924040079117, + 0.005095822270959616, + 0.004803849849849939, + 0.0026573659852147102, + 0.0054397378116846085, + 0.00243651052005589, + 0.004557468928396702, + 0.002743750810623169, + 0.004867281299084425, + 0.005210442468523979, + 0.0025989722926169634, + 0.009620052762329578, + 0.01597735844552517, + 0.005565448198467493, + 0.004174377769231796, + 0.003060506656765938, + 0.0033745532855391502, + 0.003955102525651455, + 0.0036933424416929483, + 0.007002102676779032, + 0.003636724315583706, + 0.0032112887129187584, + 0.006413940340280533, + 0.003618868300691247, + 0.003961976617574692, + 0.004491461906582117, + 0.005244656931608915, + 0.008404764346778393, + 0.004112936090677977, + 0.0030073130037635565, + 0.006155729293823242, + 0.0033691374119371176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.048635248094797134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016057809814810753 + ], + "opposite_logprob_tokens": [ + -4.141057968139648 + ], + "logprob_reward": -0.016057809814810753, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016057809814810753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023012131452560425, + 0.021797556430101395, + 0.03717264533042908, + 0.031058896332979202, + 0.027186620980501175, + 0.6258525848388672, + 0.020132772624492645, + 0.1218927651643753, + 0.07277185469865799, + 0.0191221721470356 + ], + "credit_map": [ + 0.0022490834817290306, + 0.004201439209282398, + 0.0019313162192702293, + 0.003171512857079506, + 0.0035103457048535347, + 0.0023922447580844164, + 0.0028963382355868816, + 0.002659848425537348, + 0.0029620241839438677, + 0.002893873257562518, + 0.002874404890462756, + 0.0029161563143134117, + 0.0018346409779042006, + 0.002156138652935624, + 0.002326889894902706, + 0.003833424998447299, + 0.0036022441927343607, + 0.0058718458749353886, + 0.011911967769265175, + 0.002887456212192774, + 0.002692116890102625, + 0.003606087062507868, + 0.002633090363815427, + 0.003967835567891598, + 0.0058966707438230515, + 0.003148202784359455, + 0.0029293405823409557, + 0.002607045928016305, + 0.0068442970514297485, + 0.004416785668581724, + 0.00253155711106956, + 0.0026849964633584023, + 0.005531796719878912, + 0.0030526730697602034, + 0.003020210424438119, + 0.0021307712886482477, + 0.006245533935725689, + 0.001722496235743165, + 0.003135401988402009, + 0.0023477370850741863, + 0.059732843190431595, + 0.034483373165130615, + 0.03042932040989399, + 0.06999586522579193, + 0.15535244345664978, + 0.04977593570947647, + 0.1911548376083374, + 0.03492792323231697, + 0.002535776700824499, + 0.002246304415166378, + 0.0031034457497298717, + 0.0020542084239423275, + 0.003396973479539156, + 0.0019339506980031729, + 0.002820155816152692, + 0.0020419550128281116, + 0.010954818688333035, + 0.013092877343297005, + 0.005566946696490049, + 0.024451708421111107, + 0.039756014943122864, + 0.011488162912428379, + 0.011153354309499264, + 0.005428878590464592, + 0.006406349595636129, + 0.009113328531384468, + 0.005996559280902147, + 0.016660800203680992, + 0.01280104462057352, + 0.0058771902695298195, + 0.009329568594694138, + 0.0065870098769664764, + 0.001977140549570322, + 0.0018302510725334287, + 0.002123967744410038, + 0.004615776240825653, + 0.0021490787621587515, + 0.002370203612372279, + 0.0024156104773283005, + 0.0016401425236836076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.016057809814810753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted single shot, filmed with the quality of a nature documentary, slowly advances from a low angle, deep into a damp woodland. A massive monitor lizard crouches amidst rotting leaves and dark green moss. Its body, covered in ancient, beaded scales, is a dull earthy gray, its rough skin, like weathered armor, its uneven texture revealed in the dappled light filtering through the dense canopy. Its powerful limbs are firmly rooted in the moist earth, its body motionless, like a prehistoric sculpture. The only movement comes from its long, forked yellow tongue, which rapidly extends and retracts, keenly gathering information from the earthy air. As the camera draws closer, almost touching its cold, impassive black eyes, the stark contrast between stillness and movement heightens the primal, poised tension of the hunt, filling the scene with a breathtaking sense of anticipation.", + "seed": 4812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000041/group_9/index_4517_seed_4812_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06358920251950621, + "vlm_logprob": -0.06358920251950621, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06358920251950621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1156192496418953 + ], + "opposite_logprob_tokens": [ + -2.24061918258667 + ], + "logprob_reward": -0.1156192496418953, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1156192496418953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2764740586280823, + 0.020817892625927925, + 0.0305666271597147, + 0.03042047657072544, + 0.04302376136183739, + 0.07430581748485565, + 0.10572806745767593, + 0.29020777344703674, + 0.04856450855731964, + 0.07989099621772766 + ], + "credit_map": [ + 0.034051548689603806, + 0.045524947345256805, + 0.026933934539556503, + 0.030094828456640244, + 0.022462327033281326, + 0.03974183648824692, + 0.04628105089068413, + 0.031383585184812546, + 0.00220769876614213, + 0.002634208183735609, + 0.0023906808346509933, + 0.002239863621070981, + 0.003020608564838767, + 0.002812413964420557, + 0.0036997878924012184, + 0.0018126295180991292, + 0.005375930108129978, + 0.004719858057796955, + 0.0029510834719985723, + 0.002856145380064845, + 0.0031078660394996405, + 0.003634721739217639, + 0.0024057277478277683, + 0.005515294149518013, + 0.003655520500615239, + 0.004500851035118103, + 0.0029834413435310125, + 0.003229475114494562, + 0.006480316631495953, + 0.0037780844140797853, + 0.003667375072836876, + 0.002125414554029703, + 0.006223955657333136, + 0.005095507949590683, + 0.00371477659791708, + 0.0072359004989266396, + 0.00366509729065001, + 0.006217916030436754, + 0.006204336881637573, + 0.004666269291192293, + 0.009238297119736671, + 0.008590920828282833, + 0.008721640333533287, + 0.011264992877840996, + 0.009241406805813313, + 0.010975795798003674, + 0.010577774606645107, + 0.005694989580661058, + 0.01738874986767769, + 0.013916869647800922, + 0.014736533164978027, + 0.02450537122786045, + 0.0051083993166685104, + 0.01552905049175024, + 0.007313975133001804, + 0.00722911674529314, + 0.018813597038388252, + 0.023719990625977516, + 0.03801220655441284, + 0.03577020391821861, + 0.0453086793422699, + 0.04294377937912941, + 0.052451346069574356, + 0.03318796306848526, + 0.005976514425128698, + 0.006818635854870081, + 0.006648401729762554, + 0.005368830170482397, + 0.0047158291563391685, + 0.008541902527213097, + 0.005119071342051029, + 0.005375328008085489, + 0.005858561489731073, + 0.007986567914485931, + 0.011106982827186584, + 0.012252540327608585, + 0.017360327765345573, + 0.00778890261426568, + 0.008691169321537018, + 0.008845938369631767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.1156192496418953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002830429933965206 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.002830429933965206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002830429933965206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12789854407310486, + 0.05646832659840584, + 0.0656205415725708, + 0.03885805979371071, + 0.06287694722414017, + 0.04200683906674385, + 0.30639412999153137, + 0.11459434777498245, + 0.11429718136787415, + 0.0709850862622261 + ], + "credit_map": [ + 0.016314389184117317, + 0.01729460619390011, + 0.01284490805119276, + 0.013655342161655426, + 0.011497707106173038, + 0.017005065456032753, + 0.02617676742374897, + 0.013109753839671612, + 0.008696086704730988, + 0.008202866651117802, + 0.007459952495992184, + 0.004020143765956163, + 0.004905982408672571, + 0.0070225633680820465, + 0.00817285105586052, + 0.007987880147993565, + 0.01239663828164339, + 0.0090976906940341, + 0.008459417149424553, + 0.004753940273076296, + 0.005817493423819542, + 0.008057990111410618, + 0.0052272542379796505, + 0.011810116469860077, + 0.0037780676502734423, + 0.004545285832136869, + 0.0049151103012263775, + 0.0037523244973272085, + 0.004890786949545145, + 0.005344937555491924, + 0.007228210102766752, + 0.004403335507959127, + 0.011985654942691326, + 0.010368588380515575, + 0.006083634681999683, + 0.007485749665647745, + 0.005257692653685808, + 0.00890266802161932, + 0.007129319477826357, + 0.005663634277880192, + 0.005048272665590048, + 0.005369157530367374, + 0.005245847627520561, + 0.00509819108992815, + 0.003935732413083315, + 0.004826657008379698, + 0.008468582294881344, + 0.004014396108686924, + 0.04762779176235199, + 0.03924977406859398, + 0.03391602635383606, + 0.07103649526834488, + 0.01657426916062832, + 0.04505448043346405, + 0.02835220843553543, + 0.024583080783486366, + 0.007322818040847778, + 0.010432646609842777, + 0.01843307353556156, + 0.01285314466804266, + 0.015616996213793755, + 0.016886025667190552, + 0.02030777372419834, + 0.012741868384182453, + 0.012229230254888535, + 0.012302873656153679, + 0.01273493841290474, + 0.021743766963481903, + 0.01876768097281456, + 0.017392892390489578, + 0.009637830778956413, + 0.009487966075539589, + 0.005266269668936729, + 0.0062497397884726524, + 0.006910627242177725, + 0.012923920527100563, + 0.014602099545300007, + 0.010489852167665958, + 0.00760504137724638, + 0.006937541998922825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.002830429933965206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640306372195482 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640306372195482, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640306372195482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10062079131603241, + 0.0324171744287014, + 0.0453544445335865, + 0.055076662451028824, + 0.056298866868019104, + 0.03555300086736679, + 0.4796565771102905, + 0.06623800843954086, + 0.08687455207109451, + 0.04190988838672638 + ], + "credit_map": [ + 0.012149947695434093, + 0.011778977699577808, + 0.013538363389670849, + 0.013908076100051403, + 0.00642135925590992, + 0.01321958564221859, + 0.019778134301304817, + 0.009826348163187504, + 0.004459999036043882, + 0.004766032565385103, + 0.0032434193417429924, + 0.0030831280164420605, + 0.004467591643333435, + 0.004497716668993235, + 0.004368010442703962, + 0.0035312774125486612, + 0.00478679034858942, + 0.0058030616492033005, + 0.005784471984952688, + 0.005296815652400255, + 0.006043110508471727, + 0.005664243828505278, + 0.0054526603780686855, + 0.00652328971773386, + 0.00510087376460433, + 0.008518547751009464, + 0.005421324633061886, + 0.005392430815845728, + 0.011047731153666973, + 0.007629235275089741, + 0.00817958265542984, + 0.00378693756647408, + 0.011400016956031322, + 0.010166993364691734, + 0.005958161782473326, + 0.005954086314886808, + 0.00482384767383337, + 0.00777289317920804, + 0.00508127873763442, + 0.005141590256243944, + 0.005560753867030144, + 0.004722762852907181, + 0.004701156169176102, + 0.0038455917965620756, + 0.0035124451387673616, + 0.004918175749480724, + 0.0048185549676418304, + 0.0034735575318336487, + 0.06640873849391937, + 0.06038113683462143, + 0.04982805997133255, + 0.13133350014686584, + 0.03003949299454689, + 0.07782403379678726, + 0.03291506692767143, + 0.030926553532481194, + 0.005055922083556652, + 0.0058320448733866215, + 0.007808638270944357, + 0.015059998258948326, + 0.006144825369119644, + 0.01129243429750204, + 0.00931885838508606, + 0.005725289229303598, + 0.007557824719697237, + 0.006145492196083069, + 0.009899343363940716, + 0.015407945029437542, + 0.013460248708724976, + 0.01490164827555418, + 0.011835914105176926, + 0.007666133344173431, + 0.003429441712796688, + 0.004703269340097904, + 0.004990989807993174, + 0.009243762120604515, + 0.007690104655921459, + 0.0034957537427544594, + 0.004205305594950914, + 0.004151263274252415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.007640306372195482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895562797784805 + ], + "opposite_logprob_tokens": [ + -2.57895565032959 + ], + "logprob_reward": -0.07895562797784805, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895562797784805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4113786518573761, + 0.07993165403604507, + 0.09126859903335571, + 0.05812126770615578, + 0.0693436861038208, + 0.04417704790830612, + 0.09320651739835739, + 0.0425603985786438, + 0.05302777886390686, + 0.05698440968990326 + ], + "credit_map": [ + 0.050792284309864044, + 0.0697992667555809, + 0.04154447466135025, + 0.055036887526512146, + 0.03759117051959038, + 0.05407807230949402, + 0.06657974421977997, + 0.0359567254781723, + 0.010350062511861324, + 0.01226750947535038, + 0.009265747852623463, + 0.009169171564280987, + 0.01169790979474783, + 0.010416304692626, + 0.010250798426568508, + 0.006514142267405987, + 0.01039766613394022, + 0.015620158053934574, + 0.008313044905662537, + 0.011369071900844574, + 0.01594890095293522, + 0.011679560877382755, + 0.011666941456496716, + 0.006273261271417141, + 0.0053252847865223885, + 0.008430908434092999, + 0.006302928552031517, + 0.006440101657062769, + 0.01055846642702818, + 0.008600080385804176, + 0.007658239919692278, + 0.004805255215615034, + 0.012964068911969662, + 0.011851147748529911, + 0.0073452964425086975, + 0.007052794564515352, + 0.009347612038254738, + 0.009054227732121944, + 0.004929837770760059, + 0.006798700895160437, + 0.005202068947255611, + 0.005129165016114712, + 0.0037197063211351633, + 0.0064745028503239155, + 0.008105101995170116, + 0.005874920636415482, + 0.006436196621507406, + 0.0032353876158595085, + 0.013597422279417515, + 0.012445873580873013, + 0.009718713350594044, + 0.023600194603204727, + 0.006079758983105421, + 0.014859984628856182, + 0.006088102236390114, + 0.0068164728581905365, + 0.0038454397581517696, + 0.005058916285634041, + 0.0057730767875909805, + 0.005218386184424162, + 0.005116327200084925, + 0.00604839576408267, + 0.00645493483170867, + 0.005044921301305294, + 0.0049123200587928295, + 0.00784352608025074, + 0.008166470564901829, + 0.009001517668366432, + 0.006369576323777437, + 0.006699909456074238, + 0.004919904749840498, + 0.005114552099257708, + 0.003460954176262021, + 0.005150131415575743, + 0.007092801854014397, + 0.009646010585129261, + 0.010862834751605988, + 0.007634621579200029, + 0.007561215199530125, + 0.00557584036141634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.07895562797784805 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive monitor lizard with dull earthy gray skin and beaded scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard crouching amidst rotting leaves and dark green moss in a damp woodland environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's body motionless while its long, forked yellow tongue rapidly extends and retracts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the lizard's black eyes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance from a low angle, drawing closer to the lizard?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290039867162704 + ], + "opposite_logprob_tokens": [ + -2.237900495529175 + ], + "logprob_reward": -0.11290039867162704, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290039867162704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25540101528167725, + 0.046572037041187286, + 0.06970079988241196, + 0.022120753303170204, + 0.05551408603787422, + 0.04372940957546234, + 0.3598066568374634, + 0.039143286645412445, + 0.05386762693524361, + 0.05414428934454918 + ], + "credit_map": [ + 0.032705631107091904, + 0.041506312787532806, + 0.02522081881761551, + 0.02867642417550087, + 0.02167070470750332, + 0.037657395005226135, + 0.04122451692819595, + 0.026739206165075302, + 0.006428923923522234, + 0.006897293496876955, + 0.005355849862098694, + 0.004821386653929949, + 0.005568945314735174, + 0.006561857648193836, + 0.005759551655501127, + 0.005178227089345455, + 0.014526919461786747, + 0.01048152707517147, + 0.007614060305058956, + 0.004970966372638941, + 0.005988746881484985, + 0.008103678934276104, + 0.006104856263846159, + 0.011910049244761467, + 0.0023692722897976637, + 0.002416796050965786, + 0.0032324204221367836, + 0.002323341555893421, + 0.002912670373916626, + 0.0032086961437016726, + 0.0032170466147363186, + 0.0024405112490057945, + 0.008999630808830261, + 0.008829908445477486, + 0.007049814332276583, + 0.0050633009523153305, + 0.007454507052898407, + 0.007336823735386133, + 0.004952710587531328, + 0.00582739245146513, + 0.0033768415451049805, + 0.004741926211863756, + 0.0036663771606981754, + 0.007679756730794907, + 0.006323319394141436, + 0.006607863586395979, + 0.00748791079968214, + 0.0038454150781035423, + 0.0581040158867836, + 0.05003118887543678, + 0.038804791867733, + 0.08137987554073334, + 0.023523325100541115, + 0.051904838532209396, + 0.02927722968161106, + 0.026781417429447174, + 0.003547772765159607, + 0.0036775940097868443, + 0.004975183866918087, + 0.006263976451009512, + 0.004356215242296457, + 0.0070069837383925915, + 0.005526594817638397, + 0.0037889662198722363, + 0.003503348445519805, + 0.004972805734723806, + 0.008430873975157738, + 0.01182557176798582, + 0.011512979865074158, + 0.006449768785387278, + 0.003435223363339901, + 0.003737054066732526, + 0.004076980985701084, + 0.0039289179258048534, + 0.008296852931380272, + 0.011678352952003479, + 0.008961508981883526, + 0.008736640214920044, + 0.004253199324011803, + 0.004211835563182831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.11290039867162704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_tensor([6805, 6805])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_tensor([6805, 6805])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..32b1b9e27db996bab7429d1aa8bd238ce1254de6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_tensor([6805, 6805])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up immerses the viewer in an artistic interpretation of marine life, capturing the complex skeletal structure of a blue sunray algae* rendered in the refined style of Chinese ink painting. The coral's form is a complex network of crystalline branches, defined by sharp, calligraphic lines and textures, evoking the sensation of ink on paper. The color palette is strictly limited to deep, resonant blues and velvety, bottomless blacks, the stark contrast creating a sense of depth and form. From within the countless tiny pores of this dark, branching skeleton, translucent polyps, painted in a paler, ethereal cobalt, rhythmically expand and contract. They delicately unfold, their feathery tentacles pulsating in silent, hypnotic synchrony before rapidly retreating back into the structure, a continuous, wave-like motion flowing through the entire colony. This collective action creates a powerful impression of a single living organism slowly chewing through the surrounding void, which is not water but an abstract background of bleeding, graded ink washes, establishing an atmosphere of living art that is mysterious and profoundly tranquil.", + "seed": 26031, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26031_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2101225558668375, + "vlm_logprob": -0.2101225558668375, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.2101225558668375, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6994423270225525 + ], + "opposite_logprob_tokens": [ + -0.6994423270225525 + ], + "logprob_reward": -0.6994423270225525, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6994423270225525, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1998831331729889, + 0.06186239793896675, + 0.1696416586637497, + 0.05213343724608421, + 0.07227316498756409, + 0.11755669116973877, + 0.1450691968202591, + 0.09761393815279007, + 0.05133657529950142, + 0.032629773020744324 + ], + "credit_map": [ + 0.01792917586863041, + 0.019786713644862175, + 0.038955118507146835, + 0.021897396072745323, + 0.038656625896692276, + 0.019778674468398094, + 0.02315724827349186, + 0.019722191616892815, + 0.0058248634450137615, + 0.006837218068540096, + 0.009118130430579185, + 0.011495099402964115, + 0.010968084447085857, + 0.005328143015503883, + 0.007705214899033308, + 0.00458564842119813, + 0.012664047069847584, + 0.02303306572139263, + 0.015531404875218868, + 0.04059059917926788, + 0.028925303369760513, + 0.014728142879903316, + 0.020624367520213127, + 0.013544735498726368, + 0.0035434833262115717, + 0.00513065978884697, + 0.004798375070095062, + 0.00764821982011199, + 0.008089492097496986, + 0.00497791962698102, + 0.011623784899711609, + 0.006321496795862913, + 0.008286666125059128, + 0.004288188181817532, + 0.02342534251511097, + 0.00431238766759634, + 0.018144812434911728, + 0.00425143726170063, + 0.0043052369728684425, + 0.005259091965854168, + 0.01626913994550705, + 0.011328688822686672, + 0.010275152511894703, + 0.019700413569808006, + 0.018556108698248863, + 0.01105688139796257, + 0.018444931134581566, + 0.011925376951694489, + 0.016225917264819145, + 0.03422081843018532, + 0.012500873766839504, + 0.01711275614798069, + 0.02348584681749344, + 0.012521816417574883, + 0.011623701080679893, + 0.01737748086452484, + 0.00797649472951889, + 0.028303159400820732, + 0.0055124154314398766, + 0.007879084907472134, + 0.03164637088775635, + 0.005828983150422573, + 0.0059394221752882, + 0.004528004676103592, + 0.004990157671272755, + 0.009691174142062664, + 0.005014289636164904, + 0.006922193802893162, + 0.007021685130894184, + 0.004924120847135782, + 0.008294295519590378, + 0.0044786566868424416, + 0.002699968172237277, + 0.00431549409404397, + 0.004953453317284584, + 0.006706822197884321, + 0.003707283642143011, + 0.003885312005877495, + 0.0034751908387988806, + 0.0028862503822892904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.6994423270225525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292989552021027 + ], + "opposite_logprob_tokens": [ + -2.2379298210144043 + ], + "logprob_reward": -0.11292989552021027, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292989552021027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416327804327011, + 0.08385355025529861, + 0.07365686446428299, + 0.04674903303384781, + 0.1648276150226593, + 0.17093788087368011, + 0.18106310069561005, + 0.06718893349170685, + 0.05285730957984924, + 0.017232881858944893 + ], + "credit_map": [ + 0.01537146233022213, + 0.016604077070951462, + 0.019441526383161545, + 0.020047688856720924, + 0.024561675265431404, + 0.019051695242524147, + 0.013952808454632759, + 0.012601858004927635, + 0.007148300297558308, + 0.009882590733468533, + 0.013473332859575748, + 0.010231810621917248, + 0.020377233624458313, + 0.007714504841715097, + 0.007671964354813099, + 0.007353819441050291, + 0.006539807189255953, + 0.01324264332652092, + 0.007698365021497011, + 0.010713955387473106, + 0.00954410806298256, + 0.010953360237181187, + 0.00866109412163496, + 0.00630353856831789, + 0.004460531752556562, + 0.006156067829579115, + 0.004435076843947172, + 0.006362863350659609, + 0.005752480588853359, + 0.006466119084507227, + 0.008621002547442913, + 0.004494893364608288, + 0.011455624364316463, + 0.005337102804332972, + 0.06083100289106369, + 0.0077165099792182446, + 0.06059851124882698, + 0.0068726022727787495, + 0.004722415003925562, + 0.007293847389519215, + 0.019549649208784103, + 0.017608068883419037, + 0.021594054996967316, + 0.029938889667391777, + 0.03224172443151474, + 0.016155513003468513, + 0.017201008275151253, + 0.01664898917078972, + 0.020918171852827072, + 0.03732718899846077, + 0.01914343237876892, + 0.022395413368940353, + 0.028505057096481323, + 0.0194094181060791, + 0.013740026392042637, + 0.019624412059783936, + 0.006627955008298159, + 0.00911401491612196, + 0.010120898485183716, + 0.008016281761229038, + 0.009718840941786766, + 0.006542879156768322, + 0.010764789767563343, + 0.00628327764570713, + 0.005040745250880718, + 0.006807700265198946, + 0.005552067421376705, + 0.007475930731743574, + 0.007128207944333553, + 0.005205871537327766, + 0.010495501570403576, + 0.005151285789906979, + 0.002153725828975439, + 0.0022616651840507984, + 0.0029365301597863436, + 0.0023640673607587814, + 0.001966118114069104, + 0.002359672449529171, + 0.0016872910782694817, + 0.0015038135461509228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.11292989552021027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089951619505882 + ], + "opposite_logprob_tokens": [ + -4.141089916229248 + ], + "logprob_reward": -0.016089951619505882, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089951619505882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1944543868303299, + 0.14653950929641724, + 0.055607300251722336, + 0.07432141155004501, + 0.10454032570123672, + 0.13540513813495636, + 0.10211502760648727, + 0.050466760993003845, + 0.049240000545978546, + 0.08731010556221008 + ], + "credit_map": [ + 0.020221535116434097, + 0.01373659074306488, + 0.0461680106818676, + 0.025505760684609413, + 0.02565017156302929, + 0.026183340698480606, + 0.02048894762992859, + 0.016500037163496017, + 0.009769169613718987, + 0.01245913840830326, + 0.02394598349928856, + 0.02429957129061222, + 0.03442369028925896, + 0.01408351305872202, + 0.0184257123619318, + 0.00913273449987173, + 0.005775542929768562, + 0.0052552009001374245, + 0.008755501359701157, + 0.010019585490226746, + 0.007793060503900051, + 0.005888801533728838, + 0.006173702422529459, + 0.005945906974375248, + 0.009613532572984695, + 0.011940968222916126, + 0.00723486952483654, + 0.007260752841830254, + 0.01370706781744957, + 0.011738256551325321, + 0.005221839528530836, + 0.007604122627526522, + 0.007727581076323986, + 0.006326624192297459, + 0.029908044263720512, + 0.007443699054419994, + 0.030760815367102623, + 0.007590354885905981, + 0.006250351667404175, + 0.008532860316336155, + 0.022285763174295425, + 0.012143963016569614, + 0.010197106748819351, + 0.013753109611570835, + 0.029011208564043045, + 0.015091998502612114, + 0.015903055667877197, + 0.017018938437104225, + 0.011608230881392956, + 0.022188052535057068, + 0.010502655990421772, + 0.01246021967381239, + 0.015781691297888756, + 0.009460918605327606, + 0.007367646787315607, + 0.012745622545480728, + 0.004744570702314377, + 0.010768982581794262, + 0.004095874726772308, + 0.005373556166887283, + 0.011117152869701385, + 0.005640320014208555, + 0.00392164196819067, + 0.004804666619747877, + 0.006298360880464315, + 0.008344028145074844, + 0.006843028124421835, + 0.006865881849080324, + 0.0059074824675917625, + 0.004678718280047178, + 0.005715077742934227, + 0.004587425384670496, + 0.008973220363259315, + 0.010735454969108105, + 0.010859751142561436, + 0.01157015934586525, + 0.010131174698472023, + 0.012727455236017704, + 0.009888969361782074, + 0.012423933483660221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016089951619505882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059059962630272 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.02059059962630272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059059962630272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10616261512041092, + 0.13936173915863037, + 0.049555085599422455, + 0.03402404859662056, + 0.14524608850479126, + 0.23846599459648132, + 0.12437848746776581, + 0.06874899566173553, + 0.039787162095308304, + 0.054269760847091675 + ], + "credit_map": [ + 0.011302338913083076, + 0.01098710484802723, + 0.017348378896713257, + 0.014189640991389751, + 0.016226811334490776, + 0.012341936118900776, + 0.010672704316675663, + 0.013093707151710987, + 0.009507490321993828, + 0.011973549611866474, + 0.018133271485567093, + 0.02225959114730358, + 0.045778539031744, + 0.009789724834263325, + 0.011117584072053432, + 0.010801986791193485, + 0.00558571657165885, + 0.006641053594648838, + 0.005483090411871672, + 0.006797637324780226, + 0.007898672483861446, + 0.0058913337998092175, + 0.005104247014969587, + 0.006153335329145193, + 0.0024193834979087114, + 0.003914978820830584, + 0.003579871030524373, + 0.004695407114923, + 0.0044908965937793255, + 0.0028469639364629984, + 0.008322933688759804, + 0.0037536167073994875, + 0.00952065084129572, + 0.0074410936795175076, + 0.050090242177248, + 0.007650917861610651, + 0.04877524822950363, + 0.007948325015604496, + 0.006181348580867052, + 0.0076382700353860855, + 0.031235579401254654, + 0.019133074209094048, + 0.02391969971358776, + 0.041521377861499786, + 0.0442843921482563, + 0.01959695667028427, + 0.03677280247211456, + 0.02200210466980934, + 0.011797258630394936, + 0.03429567068815231, + 0.009740084409713745, + 0.01171906478703022, + 0.02150428481400013, + 0.01134802121669054, + 0.009388172067701817, + 0.014585930854082108, + 0.005829450208693743, + 0.012186238542199135, + 0.006328694988042116, + 0.009293857961893082, + 0.018020199611783028, + 0.005716640502214432, + 0.006970424670726061, + 0.0044034854508936405, + 0.0033343974500894547, + 0.0073255859315395355, + 0.003946725279092789, + 0.005988235119730234, + 0.006696183700114489, + 0.004709640052169561, + 0.004241596441715956, + 0.00354480161331594, + 0.009592579677700996, + 0.007424452807754278, + 0.009467631578445435, + 0.00596719142049551, + 0.006413003895431757, + 0.0048944978043437, + 0.004952245391905308, + 0.005558157805353403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.02059059962630272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156000554561615 + ], + "opposite_logprob_tokens": [ + -1.7015600204467773 + ], + "logprob_reward": -0.20156000554561615, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156000554561615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2531897723674774, + 0.15357354283332825, + 0.08632928133010864, + 0.03945208340883255, + 0.05456191301345825, + 0.09995436668395996, + 0.11344146728515625, + 0.1029881089925766, + 0.02859644591808319, + 0.0679129883646965 + ], + "credit_map": [ + 0.024992523714900017, + 0.024282066151499748, + 0.04436377063393593, + 0.03211524337530136, + 0.04371305927634239, + 0.01935281790792942, + 0.029070978984236717, + 0.03529931977391243, + 0.011775792576372623, + 0.013808971270918846, + 0.02107859030365944, + 0.02855190820991993, + 0.03689783811569214, + 0.014285252429544926, + 0.016654064878821373, + 0.010521117597818375, + 0.007865460589528084, + 0.008417393080890179, + 0.013184944167733192, + 0.01845516636967659, + 0.012516863644123077, + 0.007695606444031, + 0.01039934903383255, + 0.007794504519551992, + 0.003499152371659875, + 0.004236975219100714, + 0.003735555103048682, + 0.004296148661524057, + 0.0059477463364601135, + 0.004276262130588293, + 0.0086800716817379, + 0.00478017283603549, + 0.006452028639614582, + 0.005869689397513866, + 0.008275208994746208, + 0.005951508414000273, + 0.009385868906974792, + 0.0057808212004601955, + 0.00658164219930768, + 0.0062651424668729305, + 0.01123295072466135, + 0.011525161564350128, + 0.010447840206325054, + 0.01695104129612446, + 0.01876719295978546, + 0.008258278481662273, + 0.01348027866333723, + 0.009291619062423706, + 0.013892194256186485, + 0.018817279487848282, + 0.012926008552312851, + 0.014347463846206665, + 0.016944466158747673, + 0.011946877464652061, + 0.010393005795776844, + 0.014174175448715687, + 0.008823717013001442, + 0.025074325501918793, + 0.007081799674779177, + 0.008198627270758152, + 0.03132396936416626, + 0.007452028803527355, + 0.00576711306348443, + 0.009266535751521587, + 0.0030543908942490816, + 0.004318921826779842, + 0.0028764919843524694, + 0.003541187383234501, + 0.005074010230600834, + 0.0038845704402774572, + 0.003622336545959115, + 0.002224535448476672, + 0.00805655773729086, + 0.00897391326725483, + 0.008540406823158264, + 0.010837754234671593, + 0.0074380566366016865, + 0.008527843281626701, + 0.006660121493041515, + 0.008878330700099468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.20156000554561615 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in an artistic interpretation of marine life, capturing the complex skeletal structure of a blue sunray algae* rendered in the refined style of Chinese ink painting. The coral's form is a complex network of crystalline branches, defined by sharp, calligraphic lines and textures, evoking the sensation of ink on paper. The color palette is strictly limited to deep, resonant blues and velvety, bottomless blacks, the stark contrast creating a sense of depth and form. From within the countless tiny pores of this dark, branching skeleton, translucent polyps, painted in a paler, ethereal cobalt, rhythmically expand and contract. They delicately unfold, their feathery tentacles pulsating in silent, hypnotic synchrony before rapidly retreating back into the structure, a continuous, wave-like motion flowing through the entire colony. This collective action creates a powerful impression of a single living organism slowly chewing through the surrounding void, which is not water but an abstract background of bleeding, graded ink washes, establishing an atmosphere of living art that is mysterious and profoundly tranquil.", + "seed": 26032, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26032_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24545540288090706, + "vlm_logprob": -0.24545540288090706, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.24545540288090706, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9055668711662292 + ], + "opposite_logprob_tokens": [ + -0.5305668711662292 + ], + "logprob_reward": -0.9055668711662292, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9055668711662292, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10989642888307571, + 0.047582294791936874, + 0.4723888337612152, + 0.044720374047756195, + 0.0660969465970993, + 0.06005597859621048, + 0.028753070160746574, + 0.07879083603620529, + 0.031733639538288116, + 0.059981610625982285 + ], + "credit_map": [ + 0.011786753311753273, + 0.011142885312438011, + 0.009779476560652256, + 0.013954694382846355, + 0.012348211370408535, + 0.020291078835725784, + 0.016726959496736526, + 0.01386637706309557, + 0.008436254225671291, + 0.00480129849165678, + 0.005322935059666634, + 0.005739788059145212, + 0.005047163460403681, + 0.005478258710354567, + 0.008109678514301777, + 0.004646917339414358, + 0.02881699800491333, + 0.11570743471384048, + 0.03462011367082596, + 0.02858142927289009, + 0.13123580813407898, + 0.04700908437371254, + 0.04114845395088196, + 0.04526952654123306, + 0.004608858842402697, + 0.003929222468286753, + 0.005611538887023926, + 0.004331701900810003, + 0.007689271122217178, + 0.005967254284769297, + 0.006968096364289522, + 0.00561442831531167, + 0.005548486951738596, + 0.006953047122806311, + 0.012190028093755245, + 0.00708326930180192, + 0.009441259317100048, + 0.005275634583085775, + 0.014187362045049667, + 0.0054178559221327305, + 0.003973828628659248, + 0.009079717099666595, + 0.004413723945617676, + 0.005381935276091099, + 0.004473325330764055, + 0.01447579637169838, + 0.005916404537856579, + 0.01234124880284071, + 0.0024815888609737158, + 0.006530965678393841, + 0.0033120259176939726, + 0.002789160003885627, + 0.0035273099783807993, + 0.0037421348970383406, + 0.0034867359790951014, + 0.002883146982640028, + 0.0056402054615318775, + 0.007467733230441809, + 0.014088297262787819, + 0.011144478805363178, + 0.006441520061343908, + 0.010267725214362144, + 0.01224584598094225, + 0.01149502582848072, + 0.0033758438657969236, + 0.003929174039512873, + 0.005371475592255592, + 0.0033350633457303047, + 0.005194548983126879, + 0.0032252103555947542, + 0.004114073235541582, + 0.00318825151771307, + 0.005523373372852802, + 0.006090753246098757, + 0.005829364527016878, + 0.003924789372831583, + 0.005110674072057009, + 0.01441639568656683, + 0.006936934776604176, + 0.012149323709309101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.9055668711662292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904949575662613 + ], + "opposite_logprob_tokens": [ + -2.4640495777130127 + ], + "logprob_reward": -0.08904949575662613, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904949575662613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.482021301984787, + 0.04385003075003624, + 0.08151645213365555, + 0.051818106323480606, + 0.10249766707420349, + 0.1349930614233017, + 0.026694955304265022, + 0.013661611825227737, + 0.0250898078083992, + 0.03785702586174011 + ], + "credit_map": [ + 0.028630880638957024, + 0.05043787881731987, + 0.03374999761581421, + 0.04598834738135338, + 0.14220447838306427, + 0.0548463836312294, + 0.05994957685470581, + 0.06621374189853668, + 0.006408835295587778, + 0.004548742901533842, + 0.004960956051945686, + 0.00645536370575428, + 0.005331206135451794, + 0.004079299047589302, + 0.00594501243904233, + 0.006120617967098951, + 0.008642962202429771, + 0.011956303380429745, + 0.008330007083714008, + 0.009962251409888268, + 0.011828319169580936, + 0.010627029463648796, + 0.011868766508996487, + 0.008300806395709515, + 0.005006242077797651, + 0.004772346466779709, + 0.004621161613613367, + 0.005061030853539705, + 0.012985827401280403, + 0.008536426350474358, + 0.005633856635540724, + 0.005201216321438551, + 0.009800652042031288, + 0.01016023475676775, + 0.01695692166686058, + 0.013582686893641949, + 0.018880605697631836, + 0.008854661136865616, + 0.015364067628979683, + 0.008897833526134491, + 0.007819072343409061, + 0.011912173591554165, + 0.013350510969758034, + 0.010697389021515846, + 0.025690848007798195, + 0.029945453628897667, + 0.01175350695848465, + 0.023824097588658333, + 0.0019289751071482897, + 0.0018530457746237516, + 0.0016094875754788518, + 0.0020190083887428045, + 0.0022517612669616938, + 0.008393287658691406, + 0.0026921252720057964, + 0.005947262514382601, + 0.001422942616045475, + 0.0013042079517617822, + 0.001906536053866148, + 0.0014227258507162333, + 0.0010523318778723478, + 0.0028118258342146873, + 0.0013302055886015296, + 0.002410835586488247, + 0.0020705624483525753, + 0.0037164322566241026, + 0.003069981001317501, + 0.002440560609102249, + 0.0037237766664475203, + 0.00339151150546968, + 0.0036745008546859026, + 0.003002482932060957, + 0.003797799116000533, + 0.006092608440667391, + 0.004448468331247568, + 0.004284801427274942, + 0.004862923640757799, + 0.0038115528877824545, + 0.006234389264136553, + 0.004324481822550297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.08904949575662613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373829692602158 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.026373829692602158, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373829692602158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.333749383687973, + 0.06443687528371811, + 0.22325845062732697, + 0.0736488476395607, + 0.027108585461974144, + 0.11991012096405029, + 0.04307844862341881, + 0.015352579765021801, + 0.02452045865356922, + 0.07493618875741959 + ], + "credit_map": [ + 0.01985015533864498, + 0.03853935748338699, + 0.02056434378027916, + 0.030219539999961853, + 0.09847255796194077, + 0.04002843424677849, + 0.03242957964539528, + 0.05364541709423065, + 0.007901011034846306, + 0.005058480426669121, + 0.006172906141728163, + 0.0132896201685071, + 0.007676116656512022, + 0.0038216370157897472, + 0.012412641197443008, + 0.008104470558464527, + 0.015992142260074615, + 0.05360272154211998, + 0.01732310652732849, + 0.01828775182366371, + 0.05088276043534279, + 0.026899397373199463, + 0.020107118412852287, + 0.020163467153906822, + 0.007152235135436058, + 0.0070112417452037334, + 0.009589821100234985, + 0.005929640028625727, + 0.012461919337511063, + 0.009540349245071411, + 0.01488442998379469, + 0.007079216185957193, + 0.003125069197267294, + 0.0025349559728056192, + 0.00474034296348691, + 0.00317549891769886, + 0.003427190938964486, + 0.002852371660992503, + 0.004721374716609716, + 0.0025317843537777662, + 0.00714309373870492, + 0.009669866412878036, + 0.012445575557649136, + 0.009346000850200653, + 0.022475572302937508, + 0.02607802115380764, + 0.012226203456521034, + 0.020525794476270676, + 0.0028587798587977886, + 0.001535292831249535, + 0.003134022932499647, + 0.0022074170410633087, + 0.0021875197999179363, + 0.015998782590031624, + 0.002784367185086012, + 0.012372269295156002, + 0.0018549105152487755, + 0.0014368874253705144, + 0.002020803978666663, + 0.0016265481244772673, + 0.0017138344701379538, + 0.0023327558301389217, + 0.0024388672318309546, + 0.0019279735861346126, + 0.004577449522912502, + 0.0034857983700931072, + 0.0027892731595784426, + 0.0022495281882584095, + 0.002543507842347026, + 0.003220541402697563, + 0.0033980815205723047, + 0.002256282838061452, + 0.00885655079036951, + 0.009565941989421844, + 0.010158983059227467, + 0.008445695042610168, + 0.009185876697301865, + 0.008488752879202366, + 0.011232098564505577, + 0.009002295322716236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026373829692602158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366397738456726 + ], + "opposite_logprob_tokens": [ + -3.6513664722442627 + ], + "logprob_reward": -0.026366397738456726, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366397738456726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.380944162607193, + 0.07458532601594925, + 0.17653946578502655, + 0.03187544643878937, + 0.033853594213724136, + 0.14642271399497986, + 0.01618306338787079, + 0.020020103082060814, + 0.033502742648124695, + 0.0860733613371849 + ], + "credit_map": [ + 0.02208746410906315, + 0.043468598276376724, + 0.02602113038301468, + 0.033915597945451736, + 0.11637986451387405, + 0.03808152303099632, + 0.05507997050881386, + 0.04591001570224762, + 0.006832392420619726, + 0.005933765321969986, + 0.006593295838683844, + 0.017467306926846504, + 0.011081365868449211, + 0.0047117494978010654, + 0.013643220067024231, + 0.008322233334183693, + 0.010960979387164116, + 0.05136444419622421, + 0.011414899490773678, + 0.00868906732648611, + 0.05139536038041115, + 0.012090216390788555, + 0.015888439491391182, + 0.014736057259142399, + 0.0029404100496321917, + 0.002399265766143799, + 0.0049790083430707455, + 0.0026132927741855383, + 0.006585939787328243, + 0.0030495442915707827, + 0.0069832345470786095, + 0.0023247497156262398, + 0.0032525479327887297, + 0.003517576027661562, + 0.004693064372986555, + 0.005112830549478531, + 0.006370470393449068, + 0.0038315579295158386, + 0.004010617733001709, + 0.003064929973334074, + 0.008900412358343601, + 0.0127906808629632, + 0.01518961414694786, + 0.011663149110972881, + 0.03414459154009819, + 0.025402691215276718, + 0.01917335018515587, + 0.019158219918608665, + 0.0018301364034414291, + 0.00279197096824646, + 0.0022660603281110525, + 0.001835892559029162, + 0.0020712274126708508, + 0.0016744105378165841, + 0.0022703721188008785, + 0.001442992128431797, + 0.0017496957443654537, + 0.0016828794032335281, + 0.0027604554779827595, + 0.002209003083407879, + 0.0016561989905312657, + 0.004463033750653267, + 0.002072804607450962, + 0.003426030743867159, + 0.003124485956504941, + 0.003744714893400669, + 0.0037575720343738794, + 0.002880324609577656, + 0.005967073608189821, + 0.005780378356575966, + 0.003232775954529643, + 0.005015416536480188, + 0.0075292205438017845, + 0.009783996269106865, + 0.011120802722871304, + 0.011188666336238384, + 0.0101456418633461, + 0.01331012137234211, + 0.014969277195632458, + 0.008025630377233028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026366397738456726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992042005062103 + ], + "opposite_logprob_tokens": [ + -1.8049204349517822 + ], + "logprob_reward": -0.17992042005062103, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992042005062103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628135681152344, + 0.05449497327208519, + 0.3896928131580353, + 0.07645218074321747, + 0.02724396251142025, + 0.05013361945748329, + 0.0277391504496336, + 0.02640770934522152, + 0.022863024845719337, + 0.062158990651369095 + ], + "credit_map": [ + 0.018219856545329094, + 0.029952893033623695, + 0.02094942145049572, + 0.023188890889286995, + 0.06888686865568161, + 0.027197178453207016, + 0.026632850989699364, + 0.047785621136426926, + 0.007955092005431652, + 0.006160080898553133, + 0.007074114866554737, + 0.006465479731559753, + 0.0069479807280004025, + 0.0059221056289970875, + 0.008073370903730392, + 0.005896750371903181, + 0.023678723722696304, + 0.09908758848905563, + 0.029124746099114418, + 0.02245873399078846, + 0.10322923958301544, + 0.04479140788316727, + 0.03139156848192215, + 0.0359308123588562, + 0.00808551162481308, + 0.008238555863499641, + 0.008391455747187138, + 0.006682307459414005, + 0.01503778900951147, + 0.009211781434714794, + 0.012694031931459904, + 0.00811076071113348, + 0.0031332254875451326, + 0.002545151161029935, + 0.0038411591667681932, + 0.003413210390135646, + 0.003304415848106146, + 0.0032699215225875378, + 0.004395694006234407, + 0.003341186558827758, + 0.0033854395151138306, + 0.005765901878476143, + 0.004216534085571766, + 0.004887544550001621, + 0.003337073838338256, + 0.012965545989573002, + 0.0038450458087027073, + 0.011730536818504333, + 0.002022300846874714, + 0.002122265286743641, + 0.0027101575396955013, + 0.0024069484788924456, + 0.0031199005898088217, + 0.006734667345881462, + 0.003813286079093814, + 0.0048096259124577045, + 0.0032166072633117437, + 0.0027836149092763662, + 0.0033798536751419306, + 0.0018073220271617174, + 0.0031536174938082695, + 0.005963748786598444, + 0.0027873136568814516, + 0.0033156319987028837, + 0.0019670301117002964, + 0.0030379146337509155, + 0.003148983232676983, + 0.0024402339477092028, + 0.00302918185479939, + 0.003872038098052144, + 0.0029232841916382313, + 0.002444361336529255, + 0.005266975145787001, + 0.0065805003978312016, + 0.0049409870989620686, + 0.0046697258949279785, + 0.004879343789070845, + 0.015394006855785847, + 0.006738428026437759, + 0.013689026236534119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.17992042005062103 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in an artistic interpretation of marine life, capturing the complex skeletal structure of a blue sunray algae* rendered in the refined style of Chinese ink painting. The coral's form is a complex network of crystalline branches, defined by sharp, calligraphic lines and textures, evoking the sensation of ink on paper. The color palette is strictly limited to deep, resonant blues and velvety, bottomless blacks, the stark contrast creating a sense of depth and form. From within the countless tiny pores of this dark, branching skeleton, translucent polyps, painted in a paler, ethereal cobalt, rhythmically expand and contract. They delicately unfold, their feathery tentacles pulsating in silent, hypnotic synchrony before rapidly retreating back into the structure, a continuous, wave-like motion flowing through the entire colony. This collective action creates a powerful impression of a single living organism slowly chewing through the surrounding void, which is not water but an abstract background of bleeding, graded ink washes, establishing an atmosphere of living art that is mysterious and profoundly tranquil.", + "seed": 26033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26033_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28445547856390474, + "vlm_logprob": -0.28445547856390474, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.28445547856390474, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1451472043991089 + ], + "opposite_logprob_tokens": [ + -0.3951471745967865 + ], + "logprob_reward": -1.1451472043991089, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1451472043991089, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6423337459564209, + 0.034250278025865555, + 0.01794900931417942, + 0.06487869471311569, + 0.03036373481154442, + 0.01230608019977808, + 0.07865668833255768, + 0.03414768725633621, + 0.052371930330991745, + 0.032742176204919815 + ], + "credit_map": [ + 0.056842681020498276, + 0.023206772282719612, + 0.024050304666161537, + 0.03213614970445633, + 0.035305097699165344, + 0.04022417962551117, + 0.2698509097099304, + 0.1607176661491394, + 0.0028000951278954744, + 0.0031650776509195566, + 0.002426476450636983, + 0.003930006176233292, + 0.008349284529685974, + 0.003881721058860421, + 0.005619894713163376, + 0.004077723249793053, + 0.0016201019752770662, + 0.002845674054697156, + 0.0016457897145301104, + 0.002328647766262293, + 0.0029538418166339397, + 0.0026160518173128366, + 0.0017386442050337791, + 0.0022002574987709522, + 0.0050894771702587605, + 0.0200358759611845, + 0.0032187786418944597, + 0.0035994851496070623, + 0.004176560323685408, + 0.003220487153157592, + 0.021052800118923187, + 0.004485228564590216, + 0.0023870423901826143, + 0.005603422410786152, + 0.0032327575609087944, + 0.0035192242357879877, + 0.004119222983717918, + 0.004324014764279127, + 0.0026627902407199144, + 0.004515260457992554, + 0.0009143790230154991, + 0.0015701379161328077, + 0.0016671717166900635, + 0.0014100538101047277, + 0.0014312431449070573, + 0.0014768491964787245, + 0.0012038572458550334, + 0.0026323881465941668, + 0.005075111053884029, + 0.011139296926558018, + 0.007660667411983013, + 0.010448531247675419, + 0.008058235049247742, + 0.008793383836746216, + 0.011938079260289669, + 0.015543383546173573, + 0.0027630107942968607, + 0.005199433770030737, + 0.005221334286034107, + 0.003218047320842743, + 0.003057764610275626, + 0.004275026265531778, + 0.0043856133706867695, + 0.006027456372976303, + 0.0026297408621758223, + 0.002691283356398344, + 0.0026949485763907433, + 0.0020224559120833874, + 0.003593175206333399, + 0.020254969596862793, + 0.007142600603401661, + 0.011342755518853664, + 0.002588513307273388, + 0.0032503430265933275, + 0.0028395671397447586, + 0.008250756189227104, + 0.002194391330704093, + 0.0051665841601789, + 0.004667395260185003, + 0.003784622997045517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -1.1451472043991089 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706296145915985 + ], + "opposite_logprob_tokens": [ + -2.127063035964966 + ], + "logprob_reward": -0.12706296145915985, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706296145915985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.534111738204956, + 0.054748475551605225, + 0.027531025931239128, + 0.0364772193133831, + 0.062129512429237366, + 0.008917907252907753, + 0.043127503246068954, + 0.020726019516587257, + 0.04211658984422684, + 0.17011398077011108 + ], + "credit_map": [ + 0.04633070155978203, + 0.02099166251718998, + 0.021702129393815994, + 0.028545431792736053, + 0.030642332509160042, + 0.0357985645532608, + 0.22723519802093506, + 0.12286577373743057, + 0.005856492090970278, + 0.005582096520811319, + 0.0045796288177371025, + 0.006806172430515289, + 0.006835854612290859, + 0.007805049419403076, + 0.009277635253965855, + 0.008005552925169468, + 0.0024569351226091385, + 0.004315751604735851, + 0.0020818018820136786, + 0.0035266121849417686, + 0.0043140617199242115, + 0.0036192897241562605, + 0.0031595895998179913, + 0.004056987818330526, + 0.003689677221700549, + 0.005495665594935417, + 0.0030001013074070215, + 0.00275421398691833, + 0.003288066480308771, + 0.005793637130409479, + 0.007759707048535347, + 0.004696154501289129, + 0.004644372500479221, + 0.015563960187137127, + 0.005572065711021423, + 0.007533861789852381, + 0.00826567504554987, + 0.009060966782271862, + 0.006132862064987421, + 0.005355753470212221, + 0.0007264143787324429, + 0.0016550435684621334, + 0.0010041994974017143, + 0.001027089892886579, + 0.001045397948473692, + 0.0013282380532473326, + 0.0010151972528547049, + 0.0011163277085870504, + 0.0035557965748012066, + 0.008786601945757866, + 0.004785085096955299, + 0.002776771318167448, + 0.005087126046419144, + 0.004757252521812916, + 0.006632372736930847, + 0.0067464979365468025, + 0.0017364288214594126, + 0.004866357892751694, + 0.0026072629261761904, + 0.0014702256303280592, + 0.002346598543226719, + 0.0027774786576628685, + 0.001877867616713047, + 0.00304380152374506, + 0.002290430711582303, + 0.0029009853024035692, + 0.0018339903326705098, + 0.0022666354198008776, + 0.002277932595461607, + 0.01442417036741972, + 0.0022592705208808184, + 0.013863175176084042, + 0.011186358518898487, + 0.014594754204154015, + 0.0046811457723379135, + 0.005432801321148872, + 0.007976861670613289, + 0.06452803313732147, + 0.0056694974191486835, + 0.05604454129934311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.12706296145915985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609042100608349 + ], + "opposite_logprob_tokens": [ + -4.141090393066406 + ], + "logprob_reward": -0.01609042100608349, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609042100608349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08123090118169785, + 0.07461128383874893, + 0.07735433429479599, + 0.12512566149234772, + 0.04515519738197327, + 0.029425738379359245, + 0.1365565061569214, + 0.06388168781995773, + 0.25940075516700745, + 0.10725784301757812 + ], + "credit_map": [ + 0.007266543339937925, + 0.009521350264549255, + 0.006952604744583368, + 0.009084799326956272, + 0.010456685908138752, + 0.013320036232471466, + 0.013834129087626934, + 0.010794756934046745, + 0.005595926661044359, + 0.009888778440654278, + 0.006270678713917732, + 0.0069339461624622345, + 0.021456798538565636, + 0.008935773745179176, + 0.008561640046536922, + 0.006967748980969191, + 0.007257743272930384, + 0.010465587489306927, + 0.010399258695542812, + 0.008141754195094109, + 0.010795993730425835, + 0.010826592333614826, + 0.011016079224646091, + 0.008451338857412338, + 0.013175781816244125, + 0.01585467904806137, + 0.009602018631994724, + 0.009098267182707787, + 0.012365486472845078, + 0.02077772468328476, + 0.024460136890411377, + 0.019791586324572563, + 0.006892296485602856, + 0.005224770866334438, + 0.0036884888540953398, + 0.004884088411927223, + 0.0071253529749810696, + 0.004258959088474512, + 0.007912338711321354, + 0.005168908275663853, + 0.0026059241499751806, + 0.004004194401204586, + 0.0040571014396846294, + 0.0023016082122921944, + 0.0042783538810908794, + 0.004378472920507193, + 0.0030211934354156256, + 0.004778892267495394, + 0.006746236234903336, + 0.013473915867507458, + 0.007905244827270508, + 0.0207382719963789, + 0.008306687697768211, + 0.01107565127313137, + 0.04795481264591217, + 0.020355703309178352, + 0.005575091112405062, + 0.009016203694045544, + 0.008382306434214115, + 0.004675505217164755, + 0.005372514482587576, + 0.008910102769732475, + 0.009806619957089424, + 0.01214334461838007, + 0.013201347552239895, + 0.011695901863276958, + 0.011570294387638569, + 0.007172863464802504, + 0.015148519538342953, + 0.10341502726078033, + 0.017699720337986946, + 0.07949709892272949, + 0.008304128423333168, + 0.011480110697448254, + 0.005688426550477743, + 0.01347939483821392, + 0.005872915033251047, + 0.026999158784747124, + 0.005631632637232542, + 0.029802091419696808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.01609042100608349 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498284474015236 + ], + "opposite_logprob_tokens": [ + -2.929982900619507 + ], + "logprob_reward": -0.05498284474015236, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498284474015236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2339928150177002, + 0.05045070871710777, + 0.04128270223736763, + 0.02775602787733078, + 0.16013431549072266, + 0.024177396669983864, + 0.31080612540245056, + 0.017385700717568398, + 0.09873143583536148, + 0.03528273478150368 + ], + "credit_map": [ + 0.017213312909007072, + 0.014540066942572594, + 0.012338475324213505, + 0.013594899326562881, + 0.014244819059967995, + 0.02003253437578678, + 0.09015650302171707, + 0.05187219753861427, + 0.005254511255770922, + 0.005464662332087755, + 0.0039509981870651245, + 0.006308023817837238, + 0.009442301467061043, + 0.0073268418200314045, + 0.005900724325329065, + 0.006802651099860668, + 0.003599382471293211, + 0.006630119401961565, + 0.003523186082020402, + 0.003566532162949443, + 0.006771158427000046, + 0.006017685867846012, + 0.004756538663059473, + 0.006418101955205202, + 0.002383153885602951, + 0.0035563705023378134, + 0.0026552893687039614, + 0.0025001675821840763, + 0.003558259690180421, + 0.004382176324725151, + 0.004013359546661377, + 0.004707253072410822, + 0.01848079450428486, + 0.031444039195775986, + 0.013191258534789085, + 0.019379785284399986, + 0.03248419985175133, + 0.015294798649847507, + 0.018087642267346382, + 0.011771791614592075, + 0.0020499113015830517, + 0.004134521819651127, + 0.003969501703977585, + 0.002524511655792594, + 0.0027249925769865513, + 0.003213882679119706, + 0.0026227235794067383, + 0.0029373527504503727, + 0.018230823799967766, + 0.022499138489365578, + 0.028134290128946304, + 0.015615375712513924, + 0.024209318682551384, + 0.03941737115383148, + 0.08880756795406342, + 0.07389228790998459, + 0.001500247628428042, + 0.003531333524733782, + 0.001976204803213477, + 0.0012986041838303208, + 0.002384828170761466, + 0.0022275291848927736, + 0.001938570523634553, + 0.002528383396565914, + 0.005635138135403395, + 0.00515090674161911, + 0.0054755620658397675, + 0.005124757997691631, + 0.005500205792486668, + 0.03200577571988106, + 0.007120792288333178, + 0.03271830081939697, + 0.0026963483542203903, + 0.005116215907037258, + 0.0029641748405992985, + 0.00444481335580349, + 0.0026583964936435223, + 0.008615391328930855, + 0.002731948858126998, + 0.006055449601262808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.05498284474015236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899396121501923 + ], + "opposite_logprob_tokens": [ + -2.578994035720825 + ], + "logprob_reward": -0.07899396121501923, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899396121501923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4881327450275421, + 0.037541259080171585, + 0.02884192019701004, + 0.05820508301258087, + 0.12406795471906662, + 0.028800979256629944, + 0.11988741159439087, + 0.0241793692111969, + 0.04265521466732025, + 0.047688018530607224 + ], + "credit_map": [ + 0.041459523141384125, + 0.018588341772556305, + 0.01938399113714695, + 0.024664608761668205, + 0.02741321362555027, + 0.032212287187576294, + 0.20511721074581146, + 0.11929358541965485, + 0.0040963622741401196, + 0.003630888881161809, + 0.0029140799306333065, + 0.0036357122007757425, + 0.006199661176651716, + 0.006118522025644779, + 0.005895885173231363, + 0.005050153937190771, + 0.0027282987721264362, + 0.004484399221837521, + 0.00231453450396657, + 0.0032225947361439466, + 0.00483024911954999, + 0.0038188640028238297, + 0.0032279391307383776, + 0.004215043969452381, + 0.005906788166612387, + 0.012875834479928017, + 0.004951844457536936, + 0.00434227054938674, + 0.005225237924605608, + 0.007094158791005611, + 0.008871232159435749, + 0.008937720209360123, + 0.012047738768160343, + 0.0165878776460886, + 0.011736403219401836, + 0.014870340004563332, + 0.022737490013241768, + 0.015100060030817986, + 0.018647728487849236, + 0.012340337969362736, + 0.002498446498066187, + 0.0034056473523378372, + 0.003668730380013585, + 0.0025062826462090015, + 0.002825325122103095, + 0.004828820005059242, + 0.0033064624294638634, + 0.005761269014328718, + 0.0059125698171556, + 0.00833724346011877, + 0.008204656653106213, + 0.023061126470565796, + 0.006936565041542053, + 0.009714649990200996, + 0.041858889162540436, + 0.015861717984080315, + 0.0020555327646434307, + 0.0045671421103179455, + 0.002085989573970437, + 0.002059121150523424, + 0.0024146975483745337, + 0.0038200451526790857, + 0.0034077682066708803, + 0.003769075032323599, + 0.0027287364937365055, + 0.003000334370881319, + 0.002872177166864276, + 0.0029765190556645393, + 0.0032659024000167847, + 0.013099352829158306, + 0.004270430654287338, + 0.010441770777106285, + 0.0029468040447682142, + 0.00351631548255682, + 0.0023364971857517958, + 0.005049350671470165, + 0.001990696880966425, + 0.016858741641044617, + 0.0035675843246281147, + 0.011422034353017807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.07899396121501923 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up immerses the viewer in an artistic interpretation of marine life, capturing the complex skeletal structure of a blue sunray algae* rendered in the refined style of Chinese ink painting. The coral's form is a complex network of crystalline branches, defined by sharp, calligraphic lines and textures, evoking the sensation of ink on paper. The color palette is strictly limited to deep, resonant blues and velvety, bottomless blacks, the stark contrast creating a sense of depth and form. From within the countless tiny pores of this dark, branching skeleton, translucent polyps, painted in a paler, ethereal cobalt, rhythmically expand and contract. They delicately unfold, their feathery tentacles pulsating in silent, hypnotic synchrony before rapidly retreating back into the structure, a continuous, wave-like motion flowing through the entire colony. This collective action creates a powerful impression of a single living organism slowly chewing through the surrounding void, which is not water but an abstract background of bleeding, graded ink washes, establishing an atmosphere of living art that is mysterious and profoundly tranquil.", + "seed": 26034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_1/index_6805_seed_26034_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27924498431384565, + "vlm_logprob": -0.27924498431384565, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.27924498431384565, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0632174015045166 + ], + "opposite_logprob_tokens": [ + -0.438217431306839 + ], + "logprob_reward": -1.0632174015045166, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0632174015045166, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0961337760090828, + 0.1359868049621582, + 0.10942085832357407, + 0.025336330756545067, + 0.052127860486507416, + 0.0474969856441021, + 0.06489553302526474, + 0.14783772826194763, + 0.2933955192565918, + 0.027368566021323204 + ], + "credit_map": [ + 0.011308272369205952, + 0.008988416753709316, + 0.019391650334000587, + 0.011886757798492908, + 0.008946571499109268, + 0.011186799965798855, + 0.010835508815944195, + 0.01358980592340231, + 0.017093567177653313, + 0.014739476144313812, + 0.020123310387134552, + 0.019693437963724136, + 0.008794189430773258, + 0.02104836143553257, + 0.01685520075261593, + 0.0176392775028944, + 0.021886518225073814, + 0.009819182567298412, + 0.018379757180809975, + 0.00989033468067646, + 0.00926851388067007, + 0.015132876113057137, + 0.010348141193389893, + 0.014695543795824051, + 0.00291851325891912, + 0.002605910412967205, + 0.002881208434700966, + 0.005072989966720343, + 0.003425364149734378, + 0.004240743815898895, + 0.002579584252089262, + 0.0016120185609906912, + 0.00490048760548234, + 0.01052534207701683, + 0.004998208489269018, + 0.0060572330839931965, + 0.006515313405543566, + 0.010066760703921318, + 0.0053585064597427845, + 0.0037060119211673737, + 0.007073371205478907, + 0.007228251080960035, + 0.004492618143558502, + 0.006965778302401304, + 0.00755042489618063, + 0.005379886366426945, + 0.004990899469703436, + 0.0038157599046826363, + 0.006467563100159168, + 0.013122675009071827, + 0.008249619975686073, + 0.006394220516085625, + 0.009747911244630814, + 0.006896084174513817, + 0.00888093002140522, + 0.005136533174663782, + 0.007528927177190781, + 0.05227503925561905, + 0.006846177391707897, + 0.01273019053041935, + 0.013019351288676262, + 0.008595258928835392, + 0.0071954866871237755, + 0.039647310972213745, + 0.018197257071733475, + 0.033132072538137436, + 0.01798374205827713, + 0.025990812107920647, + 0.0953112468123436, + 0.022770674899220467, + 0.013551322743296623, + 0.06645841151475906, + 0.003821219317615032, + 0.0036361198872327805, + 0.00360748334787786, + 0.0037709432654082775, + 0.004375502932816744, + 0.0030432203784585, + 0.0025886192452162504, + 0.0025254609063267708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -1.0632174015045166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812770918011665 + ], + "opposite_logprob_tokens": [ + -3.2881276607513428 + ], + "logprob_reward": -0.03812770918011665, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812770918011665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07500200718641281, + 0.10384504497051239, + 0.16149620711803436, + 0.08401349931955338, + 0.04504315182566643, + 0.03439249098300934, + 0.066513292491436, + 0.23178929090499878, + 0.17218582332134247, + 0.02571920119225979 + ], + "credit_map": [ + 0.007941002957522869, + 0.006472872570157051, + 0.012231160886585712, + 0.008813033811748028, + 0.006349409930408001, + 0.010470449924468994, + 0.010428444482386112, + 0.012295630760490894, + 0.014237234368920326, + 0.011819989420473576, + 0.018610723316669464, + 0.013862626627087593, + 0.00819710548967123, + 0.011627846397459507, + 0.01122002862393856, + 0.014269497245550156, + 0.029100624844431877, + 0.01864091493189335, + 0.02421138435602188, + 0.017253713682293892, + 0.013939878903329372, + 0.016606101766228676, + 0.02112296037375927, + 0.02062063105404377, + 0.016030315309762955, + 0.010143203660845757, + 0.008470049127936363, + 0.011561649851500988, + 0.012075974605977535, + 0.008321943692862988, + 0.009582244791090488, + 0.007828115485608578, + 0.003799159312620759, + 0.008818171918392181, + 0.003892826847732067, + 0.0029611391946673393, + 0.0027713472954928875, + 0.013848723843693733, + 0.003915273584425449, + 0.005036511458456516, + 0.004159570671617985, + 0.004454813897609711, + 0.003321640891954303, + 0.005633375141769648, + 0.003527747467160225, + 0.004690378438681364, + 0.005479446612298489, + 0.003125516464933753, + 0.007879840210080147, + 0.014049086719751358, + 0.006587131414562464, + 0.004550626501441002, + 0.00974326767027378, + 0.01203115563839674, + 0.0059769172221422195, + 0.005695265717804432, + 0.014181827194988728, + 0.08404646068811417, + 0.009811189025640488, + 0.012003021314740181, + 0.019539495930075645, + 0.011924472637474537, + 0.011101212352514267, + 0.06918161362409592, + 0.02023973874747753, + 0.022887321189045906, + 0.020789314061403275, + 0.031730856746435165, + 0.02115076780319214, + 0.022736448794603348, + 0.016974201425909996, + 0.015677180141210556, + 0.0035094288177788258, + 0.003720768028870225, + 0.0027190917171537876, + 0.003507050918415189, + 0.0047530485317111015, + 0.0029456333722919226, + 0.0027201573830097914, + 0.0018440235871821642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.03812770918011665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060764841735363 + ], + "opposite_logprob_tokens": [ + -3.8956077098846436 + ], + "logprob_reward": -0.02060764841735363, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060764841735363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728037029504776, + 0.06818632036447525, + 0.04260330647230148, + 0.03396633639931679, + 0.039389822632074356, + 0.016661610454320908, + 0.06384493410587311, + 0.14926233887672424, + 0.46643468737602234, + 0.032370299100875854 + ], + "credit_map": [ + 0.007004535757005215, + 0.008730330504477024, + 0.009168559685349464, + 0.008721588179469109, + 0.012609339319169521, + 0.013655273243784904, + 0.019902726635336876, + 0.007488009054213762, + 0.007892599329352379, + 0.006805239710956812, + 0.010350040160119534, + 0.009822777472436428, + 0.006551602389663458, + 0.010252381674945354, + 0.009229605086147785, + 0.007282071281224489, + 0.006395483389496803, + 0.005057671573013067, + 0.005882509984076023, + 0.0063603222370147705, + 0.00407562218606472, + 0.004343496635556221, + 0.005712883081287146, + 0.004775314126163721, + 0.0046956646256148815, + 0.0034560298081487417, + 0.004836466163396835, + 0.0047918143682181835, + 0.003699235850945115, + 0.0051042186096310616, + 0.004295182414352894, + 0.0030877208337187767, + 0.004568993579596281, + 0.006295109633356333, + 0.003948693163692951, + 0.004855853971093893, + 0.004241407848894596, + 0.00720226252451539, + 0.005279525648802519, + 0.0029979695100337267, + 0.0025881603360176086, + 0.002638560254126787, + 0.001923060743138194, + 0.0013369017979130149, + 0.003342336043715477, + 0.001512271584942937, + 0.0018096542917191982, + 0.001510664471425116, + 0.005571105517446995, + 0.006357639096677303, + 0.01442346628755331, + 0.009540865197777748, + 0.0079118637368083, + 0.008315732702612877, + 0.007082920987159014, + 0.004641337785869837, + 0.008062967099249363, + 0.05361369997262955, + 0.006137419492006302, + 0.009858339093625546, + 0.011350679211318493, + 0.006904409732669592, + 0.0071311574429273605, + 0.04620366171002388, + 0.0388018973171711, + 0.05262669548392296, + 0.03814609721302986, + 0.04442497715353966, + 0.1260455846786499, + 0.044961147010326385, + 0.0278426893055439, + 0.09358558058738708, + 0.002724369987845421, + 0.005140797235071659, + 0.002209570026025176, + 0.007607894483953714, + 0.005129980854690075, + 0.0023602598812431097, + 0.002739644842222333, + 0.004457778763025999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02060764841735363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868123307824135 + ], + "opposite_logprob_tokens": [ + -3.0486812591552734 + ], + "logprob_reward": -0.04868123307824135, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868123307824135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09588048607110977, + 0.10711431503295898, + 0.16760371625423431, + 0.08960088342428207, + 0.07620149105787277, + 0.046358272433280945, + 0.035521700978279114, + 0.09778887778520584, + 0.25022992491722107, + 0.03370033577084541 + ], + "credit_map": [ + 0.012748355977237225, + 0.008734269998967648, + 0.015851404517889023, + 0.009815618395805359, + 0.01170742604881525, + 0.010348356328904629, + 0.013374646194279194, + 0.013300400227308273, + 0.011320251040160656, + 0.015898095443844795, + 0.012929628603160381, + 0.01965309865772724, + 0.009413478896021843, + 0.014456060715019703, + 0.012075560167431831, + 0.011368141509592533, + 0.034340936690568924, + 0.018430586904287338, + 0.026574494317173958, + 0.016095774248242378, + 0.014717288315296173, + 0.01831052266061306, + 0.01764499954879284, + 0.021489109843969345, + 0.012593415565788746, + 0.010176586918532848, + 0.011129599064588547, + 0.011990639381110668, + 0.012674148194491863, + 0.010503035970032215, + 0.011038318276405334, + 0.009495142847299576, + 0.00652898196130991, + 0.00836657639592886, + 0.009318457916378975, + 0.00817797239869833, + 0.006569857709109783, + 0.013807637616991997, + 0.008785976096987724, + 0.014646032825112343, + 0.006735650822520256, + 0.007388500031083822, + 0.0050092763267457485, + 0.005668038502335548, + 0.008795516565442085, + 0.004812831990420818, + 0.0036909037735313177, + 0.004257554188370705, + 0.0029547694139182568, + 0.0037809505593031645, + 0.00752272130921483, + 0.005554556380957365, + 0.002934408374130726, + 0.00587585149332881, + 0.004159683361649513, + 0.0027387600857764482, + 0.006689395755529404, + 0.03154027462005615, + 0.005085425917059183, + 0.009906621649861336, + 0.00896506942808628, + 0.004770697560161352, + 0.004607509355992079, + 0.026223886758089066, + 0.030009187757968903, + 0.038071099668741226, + 0.03821909800171852, + 0.03323928266763687, + 0.02386416681110859, + 0.03848668932914734, + 0.023341815918684006, + 0.02499859407544136, + 0.004526177421212196, + 0.004229582846164703, + 0.003923119977116585, + 0.004755248781293631, + 0.005021220073103905, + 0.004571612924337387, + 0.003281959565356374, + 0.0033914113882929087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.04868123307824135 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a blue sunray algae rendered in a Chinese ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing an extreme close-up of the coral's skeletal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the colors strictly limited to deep blues, velvety blacks, and pale cobalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent polyps with feathery tentacles rhythmically expand and contract from the pores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an abstract composition of bleeding, graded ink washes rather than water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559092938899994 + ], + "opposite_logprob_tokens": [ + -1.6005909442901611 + ], + "logprob_reward": -0.22559092938899994, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559092938899994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04797248914837837, + 0.07849151641130447, + 0.07285567373037338, + 0.014968116767704487, + 0.03043769672513008, + 0.02247224748134613, + 0.06791605055332184, + 0.12807102501392365, + 0.5220962166786194, + 0.014718983322381973 + ], + "credit_map": [ + 0.005527738016098738, + 0.004402898717671633, + 0.008669313043355942, + 0.005499611608684063, + 0.004177012946456671, + 0.007950317114591599, + 0.004791181534528732, + 0.006954413838684559, + 0.008197562769055367, + 0.008843866176903248, + 0.010747046209871769, + 0.013852737843990326, + 0.007132194936275482, + 0.00919471774250269, + 0.007898394018411636, + 0.012624994851648808, + 0.015429257415235043, + 0.006965140346437693, + 0.011939596384763718, + 0.006252356804907322, + 0.005482643377035856, + 0.00970553420484066, + 0.007772034499794245, + 0.00930911023169756, + 0.0025937133468687534, + 0.0016435796860605478, + 0.0016335449181497097, + 0.0023802497889846563, + 0.002164127305150032, + 0.0014441900420933962, + 0.0016613665502518415, + 0.0014473453629761934, + 0.004250030964612961, + 0.005698246881365776, + 0.0036446237936615944, + 0.0038105258718132973, + 0.003229378256946802, + 0.004244763869792223, + 0.0030128839425742626, + 0.0025472429115325212, + 0.0038335274439305067, + 0.0037871140521019697, + 0.002368647838011384, + 0.0023008438292890787, + 0.004100692458450794, + 0.002190524246543646, + 0.0020126926247030497, + 0.0018782052211463451, + 0.006843020673841238, + 0.012450236827135086, + 0.012691957876086235, + 0.008253193460404873, + 0.009266161359846592, + 0.005756273400038481, + 0.007870503701269627, + 0.004784703254699707, + 0.007010955363512039, + 0.045243486762046814, + 0.005602599587291479, + 0.009664401412010193, + 0.010774023830890656, + 0.006057410500943661, + 0.00607963465154171, + 0.03763850778341293, + 0.04221075773239136, + 0.05611634626984596, + 0.03846455737948418, + 0.047777652740478516, + 0.15217190980911255, + 0.0456722192466259, + 0.02857360430061817, + 0.1111091673374176, + 0.001891868538223207, + 0.002093857852742076, + 0.001613106345757842, + 0.0021364002022892237, + 0.0027202507480978966, + 0.0015858081169426441, + 0.0015025308821350336, + 0.0011751609854400158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.22559092938899994 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_tensor([5824, 5824])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_tensor([5824, 5824])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..951de2a6bc5a18b1598242cf5820f70ed109b3b0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_tensor([5824, 5824])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, shot at eye level, captures the intimate moment of reunion in a sun-drenched kitchen. An older woman with soft, silver-white hair pulled back in a loose bun and a young man in his twenties recline at a rustic wooden table, embraced in a warm, gentle embrace. The woman, wearing a simple cream-colored blouse and a faded floral apron, has her eyes closed, a serene smile on her face, her wrinkled hands resting tenderly on the man's shoulders. The man, wearing a navy blue Henley shirt, wraps his arms firmly around her back, his face resting lightly against her temples, a look of relief and affection. The centerpiece of the scene is a freshly baked pie on a white ceramic plate, nestled directly between them; its golden-brown lattice crust gleams, a delicate plume of steam rising from its center. Golden afternoon sunlight pours through a nearby window, illuminating the scene, capturing the dancing dust particles in the air and casting soft, warm highlights on their faces and the polished table. The backdrop showcases a cozy, live-in kitchen with white cabinets and tile countertops, creating an atmosphere of deep tranquility, comfort, and heartfelt joy.", + "seed": 9809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_5824_seed_9809_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02819935157895088, + "vlm_logprob": -0.02819935157895088, + "vlm_yes_no_margin": 3.949999952316284, + "vlm_token_credit": -0.02819935157895088, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05607986822724342 + ], + "opposite_logprob_tokens": [ + -4.056079864501953 + ], + "logprob_reward": -0.05607986822724342, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.05607986822724342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2669927477836609, + 0.08855891227722168, + 0.05478042736649513, + 0.08386975526809692, + 0.06362152844667435, + 0.04057871550321579, + 0.10424048453569412, + 0.15712358057498932, + 0.05284816771745682, + 0.0873856469988823 + ], + "credit_map": [ + 0.04338369518518448, + 0.030560726299881935, + 0.03943547606468201, + 0.05050487071275711, + 0.025246262550354004, + 0.017717497423291206, + 0.021744167432188988, + 0.03840004652738571, + 0.014788642525672913, + 0.013107245787978172, + 0.012210945598781109, + 0.013676393777132034, + 0.008105205371975899, + 0.0069710686802864075, + 0.006712465081363916, + 0.012986942194402218, + 0.009807057678699493, + 0.0066480753012001514, + 0.007053082343190908, + 0.006725061219185591, + 0.004443417303264141, + 0.007275966927409172, + 0.0035679133143275976, + 0.009259852580726147, + 0.01454941462725401, + 0.018402162939310074, + 0.007295580580830574, + 0.01208492275327444, + 0.006286201532930136, + 0.010368856601417065, + 0.007421915885061026, + 0.0074607012793421745, + 0.009118689224123955, + 0.006474078632891178, + 0.00789383240044117, + 0.011452080681920052, + 0.005941765382885933, + 0.005294866859912872, + 0.006935412064194679, + 0.010510806925594807, + 0.00676506944000721, + 0.004704347345978022, + 0.007337067276239395, + 0.004639777820557356, + 0.004876676481217146, + 0.004614450037479401, + 0.004070185590535402, + 0.003571139182895422, + 0.018845433369278908, + 0.02681182324886322, + 0.008753065951168537, + 0.012727862223982811, + 0.008480303920805454, + 0.012666014954447746, + 0.008139438927173615, + 0.007816542871296406, + 0.03214126080274582, + 0.04610101878643036, + 0.012676200829446316, + 0.020604781806468964, + 0.010260562412440777, + 0.012791421264410019, + 0.013585535809397697, + 0.008962801657617092, + 0.008679690770804882, + 0.005763871129602194, + 0.008591308258473873, + 0.009480447508394718, + 0.005548701621592045, + 0.0037244344130158424, + 0.004838218446820974, + 0.006221499294042587, + 0.017553728073835373, + 0.021300621330738068, + 0.013266352005302906, + 0.010703008621931076, + 0.006556251551955938, + 0.0064646098762750626, + 0.0064697666093707085, + 0.005071310326457024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.05607986822724342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559603899717331 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.012559603899717331, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559603899717331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24893569946289062, + 0.1479191780090332, + 0.09887096285820007, + 0.06563302129507065, + 0.048017144203186035, + 0.0442039780318737, + 0.037181951105594635, + 0.14885589480400085, + 0.056161005049943924, + 0.10422118753194809 + ], + "credit_map": [ + 0.04241978004574776, + 0.037556447088718414, + 0.027249418199062347, + 0.030908605083823204, + 0.02734595350921154, + 0.030469726771116257, + 0.026293430477380753, + 0.026692327111959457, + 0.024078963324427605, + 0.015711497515439987, + 0.016187578439712524, + 0.02124720625579357, + 0.013049447908997536, + 0.020247390493750572, + 0.018884753808379173, + 0.018512336537241936, + 0.016143620014190674, + 0.010953047312796116, + 0.008957726880908012, + 0.011991012841463089, + 0.008517173118889332, + 0.01340175699442625, + 0.014135345816612244, + 0.01477128081023693, + 0.00982202123850584, + 0.00992092676460743, + 0.005432445555925369, + 0.010828490369021893, + 0.005155941005796194, + 0.00973490346223116, + 0.007471856195479631, + 0.007266437169164419, + 0.00749785965308547, + 0.00534046022221446, + 0.0052197761833667755, + 0.007438307628035545, + 0.0043025813065469265, + 0.0047404468059539795, + 0.007378131151199341, + 0.0060995835810899734, + 0.008969666436314583, + 0.008285767398774624, + 0.004523141775280237, + 0.007379225920885801, + 0.003422777634114027, + 0.004129043780267239, + 0.002536290092393756, + 0.004958068951964378, + 0.0057865628041327, + 0.003373821498826146, + 0.0037905320059508085, + 0.006890310440212488, + 0.003427480347454548, + 0.0034464222844690084, + 0.0051888045854866505, + 0.005278018768876791, + 0.042039934545755386, + 0.054727375507354736, + 0.01032214891165495, + 0.009285963140428066, + 0.008303971029818058, + 0.014722902327775955, + 0.004549266304820776, + 0.004904336761683226, + 0.010304710827767849, + 0.013544981367886066, + 0.0044587342999875546, + 0.007583714555948973, + 0.004221804905682802, + 0.004724923521280289, + 0.006491097155958414, + 0.0048310412093997, + 0.02046893909573555, + 0.01907798834145069, + 0.007584888022392988, + 0.019997214898467064, + 0.006489102728664875, + 0.0064871846698224545, + 0.014037116430699825, + 0.010078754276037216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012559603899717331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2102704495191574, + 0.26162269711494446, + 0.11568544059991837, + 0.0779610276222229, + 0.04138544201850891, + 0.019867680966854095, + 0.03851621225476265, + 0.17023465037345886, + 0.01979806274175644, + 0.04465832561254501 + ], + "credit_map": [ + 0.03074663132429123, + 0.04667386785149574, + 0.029689379036426544, + 0.022978832945227623, + 0.020318519324064255, + 0.025861212983727455, + 0.015076294541358948, + 0.018925689160823822, + 0.03594592958688736, + 0.04882431775331497, + 0.03258616104722023, + 0.04856964200735092, + 0.019929371774196625, + 0.026440951973199844, + 0.022097498178482056, + 0.0272288229316473, + 0.01573232375085354, + 0.011344421654939651, + 0.013269064947962761, + 0.02647511474788189, + 0.009540963917970657, + 0.013083163648843765, + 0.011252696625888348, + 0.014987686648964882, + 0.011700190603733063, + 0.007444945629686117, + 0.008187045343220234, + 0.016408652067184448, + 0.006873046513646841, + 0.009767637588083744, + 0.007935838773846626, + 0.009643662720918655, + 0.006344871129840612, + 0.004255214706063271, + 0.00581733463332057, + 0.00805782712996006, + 0.0038183724973350763, + 0.003790243761613965, + 0.0047853789292275906, + 0.00451619504019618, + 0.002827754709869623, + 0.002196951536461711, + 0.002701869234442711, + 0.003606748068705201, + 0.0020525529980659485, + 0.002189249498769641, + 0.0018560558091849089, + 0.002436497015878558, + 0.004457039758563042, + 0.0064132679253816605, + 0.00333805731497705, + 0.006090381648391485, + 0.0038913912139832973, + 0.005377563647925854, + 0.004556542728096247, + 0.004391961265355349, + 0.048732854425907135, + 0.06522101163864136, + 0.011576616205275059, + 0.009508775547146797, + 0.009288917295634747, + 0.01711352914571762, + 0.0041946289129555225, + 0.00459830230101943, + 0.0033462990541011095, + 0.004170908126980066, + 0.002866544295102358, + 0.0022644340060651302, + 0.0014973898651078343, + 0.0020624345634132624, + 0.00193474639672786, + 0.0016553045716136694, + 0.008611957542598248, + 0.008629847317934036, + 0.00406512338668108, + 0.007190754637122154, + 0.0028751944191753864, + 0.0028348658233880997, + 0.006245845928788185, + 0.0042047323659062386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810384124517441 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.03810384124517441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810384124517441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.316741406917572, + 0.07676130533218384, + 0.07086792588233948, + 0.07501158118247986, + 0.04392758011817932, + 0.05333109572529793, + 0.06309281289577484, + 0.2405606359243393, + 0.02295561321079731, + 0.03674998879432678 + ], + "credit_map": [ + 0.0635879784822464, + 0.08600987493991852, + 0.037263184785842896, + 0.03571736440062523, + 0.028480853885412216, + 0.029052291065454483, + 0.01685723289847374, + 0.019772645086050034, + 0.013621502555906773, + 0.0113095473498106, + 0.012737126089632511, + 0.012952393852174282, + 0.007936687208712101, + 0.00558116938918829, + 0.004920508246868849, + 0.007702369708567858, + 0.01102964673191309, + 0.011386996135115623, + 0.011049309745430946, + 0.009895222261548042, + 0.006705086678266525, + 0.006645454093813896, + 0.004589837975800037, + 0.009566371329128742, + 0.01495592761784792, + 0.020837008953094482, + 0.007127692457288504, + 0.007832658477127552, + 0.005385673139244318, + 0.00841875746846199, + 0.0044388058595359325, + 0.0060150581412017345, + 0.00637199729681015, + 0.004151073284447193, + 0.0061521874740719795, + 0.007357789669185877, + 0.004018639214336872, + 0.0032926485873758793, + 0.0060889809392392635, + 0.006494258530437946, + 0.006723541300743818, + 0.006531625986099243, + 0.010923338122665882, + 0.01170843094587326, + 0.00424371100962162, + 0.003742791712284088, + 0.004288137890398502, + 0.0051695192232728004, + 0.011051871813833714, + 0.01745801605284214, + 0.0055731129832565784, + 0.00796046294271946, + 0.005176503676921129, + 0.006626535672694445, + 0.005187676288187504, + 0.0040586343966424465, + 0.06790915131568909, + 0.092140331864357, + 0.01620354875922203, + 0.013323881663382053, + 0.013040115125477314, + 0.024133337661623955, + 0.006701051257550716, + 0.007109223399311304, + 0.003732421202585101, + 0.00305974786169827, + 0.003153503406792879, + 0.0035706651397049427, + 0.0021375108044594526, + 0.0018378094537183642, + 0.0023815513122826815, + 0.0030824055429548025, + 0.006612391676753759, + 0.008353891782462597, + 0.006487569771707058, + 0.004350424278527498, + 0.0022556062322109938, + 0.0029010206926614046, + 0.0029170536436140537, + 0.0028720309492200613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03810384124517441 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35699939727783203, + 0.06640704721212387, + 0.05011175572872162, + 0.0757518857717514, + 0.06799954175949097, + 0.11177939176559448, + 0.05020969361066818, + 0.13920500874519348, + 0.029750073328614235, + 0.05178619921207428 + ], + "credit_map": [ + 0.07320061326026917, + 0.08189024776220322, + 0.04914572462439537, + 0.03403012827038765, + 0.03738690912723541, + 0.02661542221903801, + 0.02774740196764469, + 0.026982946321368217, + 0.010971424169838428, + 0.007237214129418135, + 0.011844269931316376, + 0.012425614520907402, + 0.006243911571800709, + 0.00517805153504014, + 0.00568473432213068, + 0.006821827031672001, + 0.007718984037637711, + 0.00516012916341424, + 0.008468314073979855, + 0.009353209286928177, + 0.005347771570086479, + 0.0037328777834773064, + 0.004660188220441341, + 0.005670277867466211, + 0.013688605278730392, + 0.01731448993086815, + 0.0069265468046069145, + 0.011769264005124569, + 0.005490904673933983, + 0.008695087395608425, + 0.006128842011094093, + 0.005738149397075176, + 0.010775420814752579, + 0.01188464742153883, + 0.008585721254348755, + 0.014011326245963573, + 0.006737567950040102, + 0.0048661306500434875, + 0.0056954496540129185, + 0.005443268921226263, + 0.028321757912635803, + 0.033994484692811966, + 0.008973656222224236, + 0.012236875481903553, + 0.006760456599295139, + 0.009190195240080357, + 0.004310496151447296, + 0.007991468533873558, + 0.010720056481659412, + 0.013757968321442604, + 0.005351227708160877, + 0.0065901088528335094, + 0.0036100184079259634, + 0.0038635018281638622, + 0.002150125801563263, + 0.0041666883043944836, + 0.036298464983701706, + 0.051318805664777756, + 0.01086411066353321, + 0.008988672867417336, + 0.008229522965848446, + 0.013536038808524609, + 0.005332044791430235, + 0.004637349396944046, + 0.005825650878250599, + 0.004563267808407545, + 0.003352784551680088, + 0.005085234064608812, + 0.002573480596765876, + 0.002433867659419775, + 0.002637208206579089, + 0.0032785814255476, + 0.00901000015437603, + 0.010257046669721603, + 0.0075667425990104675, + 0.005245435982942581, + 0.0043772142380476, + 0.0031132737640291452, + 0.006580344866961241, + 0.005636143498122692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.016066256910562515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot at eye level, captures the intimate moment of reunion in a sun-drenched kitchen. An older woman with soft, silver-white hair pulled back in a loose bun and a young man in his twenties recline at a rustic wooden table, embraced in a warm, gentle embrace. The woman, wearing a simple cream-colored blouse and a faded floral apron, has her eyes closed, a serene smile on her face, her wrinkled hands resting tenderly on the man's shoulders. The man, wearing a navy blue Henley shirt, wraps his arms firmly around her back, his face resting lightly against her temples, a look of relief and affection. The centerpiece of the scene is a freshly baked pie on a white ceramic plate, nestled directly between them; its golden-brown lattice crust gleams, a delicate plume of steam rising from its center. Golden afternoon sunlight pours through a nearby window, illuminating the scene, capturing the dancing dust particles in the air and casting soft, warm highlights on their faces and the polished table. The backdrop showcases a cozy, live-in kitchen with white cabinets and tile countertops, creating an atmosphere of deep tranquility, comfort, and heartfelt joy.", + "seed": 9810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_5824_seed_9810_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026079014502465726, + "vlm_logprob": -0.026079014502465726, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.026079014502465726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04468613117933273 + ], + "opposite_logprob_tokens": [ + -4.544686317443848 + ], + "logprob_reward": -0.04468613117933273, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.04468613117933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22769998013973236, + 0.09022451937198639, + 0.11829341948032379, + 0.06574592739343643, + 0.060144081711769104, + 0.04563312605023384, + 0.07264884561300278, + 0.11737322807312012, + 0.11180994659662247, + 0.09042689204216003 + ], + "credit_map": [ + 0.029609860852360725, + 0.02310914918780327, + 0.021009689196944237, + 0.028970440849661827, + 0.03262658044695854, + 0.030828339979052544, + 0.024324791505932808, + 0.03722112998366356, + 0.013372419402003288, + 0.008977111428976059, + 0.010073036886751652, + 0.011731269769370556, + 0.013874027878046036, + 0.010651931166648865, + 0.00830029882490635, + 0.013244427740573883, + 0.02885681577026844, + 0.010654113255441189, + 0.014051925390958786, + 0.011844371445477009, + 0.009299800731241703, + 0.009276573546230793, + 0.007794508710503578, + 0.026515314355492592, + 0.007200573105365038, + 0.005922266747802496, + 0.007553603034466505, + 0.009109169244766235, + 0.008657997474074364, + 0.008214721456170082, + 0.01058902870863676, + 0.008498570881783962, + 0.0070852916687726974, + 0.00936280470341444, + 0.0049313949421048164, + 0.007024646271020174, + 0.007195496931672096, + 0.0052338880486786366, + 0.00953739881515503, + 0.009773156605660915, + 0.008127713575959206, + 0.005865804385393858, + 0.005469196476042271, + 0.004491831175982952, + 0.006286522373557091, + 0.0045821149833500385, + 0.004113067872822285, + 0.006696871481835842, + 0.010086623951792717, + 0.007935581728816032, + 0.006772455759346485, + 0.007518950384110212, + 0.009655189700424671, + 0.006236357148736715, + 0.006317980121821165, + 0.018125709146261215, + 0.017734861001372337, + 0.013708117417991161, + 0.016383551061153412, + 0.01470747496932745, + 0.01857689768075943, + 0.012184424325823784, + 0.011422686278820038, + 0.012655207887291908, + 0.022108091041445732, + 0.008964929729700089, + 0.019463244825601578, + 0.011224078014492989, + 0.010365542955696583, + 0.009600658901035786, + 0.008494709618389606, + 0.021588686853647232, + 0.0155678391456604, + 0.006628564093261957, + 0.017666786909103394, + 0.010084813460707664, + 0.008259227499365807, + 0.006005610339343548, + 0.01050639245659113, + 0.01570766232907772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.04468613117933273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211989007890224 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014211989007890224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211989007890224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37509050965309143, + 0.08887815475463867, + 0.0905718207359314, + 0.07418558746576309, + 0.05295475572347641, + 0.03742264583706856, + 0.056257668882608414, + 0.07511831820011139, + 0.05586537346243858, + 0.09365516901016235 + ], + "credit_map": [ + 0.04793217033147812, + 0.037038177251815796, + 0.048372864723205566, + 0.04416999593377113, + 0.05191347002983093, + 0.0415053628385067, + 0.05257255584001541, + 0.051585931330919266, + 0.009817964397370815, + 0.010555295273661613, + 0.005979519337415695, + 0.010542330332100391, + 0.011407456360757351, + 0.012573858723044395, + 0.01614420861005783, + 0.01185752172023058, + 0.010934421792626381, + 0.008152710273861885, + 0.009890665300190449, + 0.009660772979259491, + 0.01187064964324236, + 0.014247260056436062, + 0.01117998268455267, + 0.01463535986840725, + 0.0071321208961308, + 0.008071820251643658, + 0.005863688886165619, + 0.0093690762296319, + 0.008169985376298428, + 0.012412368319928646, + 0.012860389426350594, + 0.010306138545274734, + 0.006854511797428131, + 0.0060981870628893375, + 0.004905452486127615, + 0.008033954538404942, + 0.006610095500946045, + 0.006878325715661049, + 0.0062318830750882626, + 0.007342347409576178, + 0.004070764407515526, + 0.004582664463669062, + 0.0031133282463997602, + 0.005641851108521223, + 0.006134362891316414, + 0.00524418568238616, + 0.004362974315881729, + 0.004272512160241604, + 0.00716982688754797, + 0.006728966720402241, + 0.007473638281226158, + 0.00612899474799633, + 0.005318252369761467, + 0.006739339325577021, + 0.010611200705170631, + 0.0060874479822814465, + 0.015094836242496967, + 0.007850736379623413, + 0.010811355896294117, + 0.00911786500364542, + 0.009985406883060932, + 0.007307555992156267, + 0.0066359215416014194, + 0.008314646780490875, + 0.005867570638656616, + 0.006408431567251682, + 0.008595953695476055, + 0.0069479080848395824, + 0.006981695536524057, + 0.0050999317318201065, + 0.007952399551868439, + 0.008011482656002045, + 0.021944252774119377, + 0.008560367859899998, + 0.011490673758089542, + 0.007966786623001099, + 0.00570429814979434, + 0.007505319081246853, + 0.006806474644690752, + 0.023676997050642967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.014211989007890224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31193408370018005, + 0.08574070781469345, + 0.07710900902748108, + 0.14895178377628326, + 0.030720258131623268, + 0.015379929915070534, + 0.05967021733522415, + 0.11161299794912338, + 0.0940084159374237, + 0.0648726373910904 + ], + "credit_map": [ + 0.027531148865818977, + 0.028142355382442474, + 0.030212944373488426, + 0.039568375796079636, + 0.0540187805891037, + 0.04603550583124161, + 0.04707564413547516, + 0.03934931755065918, + 0.009755532257258892, + 0.008632330223917961, + 0.008376515470445156, + 0.01136411540210247, + 0.014238378033041954, + 0.010652742348611355, + 0.010819370858371258, + 0.011901716701686382, + 0.014157095924019814, + 0.00658373162150383, + 0.008131994865834713, + 0.007417148910462856, + 0.007160022389143705, + 0.009412189014256, + 0.00842706486582756, + 0.01581975445151329, + 0.00963273923844099, + 0.010571223683655262, + 0.010281817056238651, + 0.02274719811975956, + 0.013306391425430775, + 0.020482903346419334, + 0.0417940653860569, + 0.02013544552028179, + 0.0036523903254419565, + 0.003428288735449314, + 0.003386778524145484, + 0.0041472758166491985, + 0.0038409745320677757, + 0.00424065301194787, + 0.004342103376984596, + 0.0036817940417677164, + 0.0016989376163110137, + 0.002375101437792182, + 0.0019132262095808983, + 0.0017767830286175013, + 0.0023742937482893467, + 0.001988612348213792, + 0.0014668222283944488, + 0.0017861532978713512, + 0.005088122561573982, + 0.005578850861638784, + 0.006669256836175919, + 0.007618667092174292, + 0.005377879366278648, + 0.0053114937618374825, + 0.016648108139634132, + 0.007377835921943188, + 0.020153868943452835, + 0.012866223230957985, + 0.018152976408600807, + 0.012305101379752159, + 0.013501770794391632, + 0.009545926004648209, + 0.01040766853839159, + 0.01467946358025074, + 0.017026938498020172, + 0.010248120874166489, + 0.014646582305431366, + 0.010439069010317326, + 0.013070784509181976, + 0.010113009251654148, + 0.007764124311506748, + 0.010699779726564884, + 0.010754442773759365, + 0.005379344802349806, + 0.005727771669626236, + 0.005829488392919302, + 0.0066486322320997715, + 0.004747106693685055, + 0.011407892219722271, + 0.014377962797880173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.029792767018079758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089267909526825 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.038089267909526825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089267909526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3177051246166229, + 0.11065923422574997, + 0.08375977724790573, + 0.10632430762052536, + 0.07346855849027634, + 0.0758839026093483, + 0.03789385408163071, + 0.05427102372050285, + 0.058687467128038406, + 0.08134674280881882 + ], + "credit_map": [ + 0.04025329649448395, + 0.040182773023843765, + 0.037441059947013855, + 0.03677283227443695, + 0.03886759281158447, + 0.06625307351350784, + 0.02494991570711136, + 0.032984595745801926, + 0.01584991067647934, + 0.0130595862865448, + 0.017125189304351807, + 0.011660142801702023, + 0.011265261098742485, + 0.012909233570098877, + 0.010242650285363197, + 0.018547257408499718, + 0.012733878567814827, + 0.00938078947365284, + 0.009547803550958633, + 0.012397238984704018, + 0.009793052449822426, + 0.008941622450947762, + 0.007492814678698778, + 0.013472573831677437, + 0.0149925472214818, + 0.013055185787379742, + 0.014437548816204071, + 0.012583120726048946, + 0.010561787523329258, + 0.01693219691514969, + 0.010033177211880684, + 0.013728746213018894, + 0.009502691216766834, + 0.008695428259670734, + 0.007570344023406506, + 0.008494862355291843, + 0.010033614002168179, + 0.008564594201743603, + 0.009888432919979095, + 0.010718593373894691, + 0.009467597119510174, + 0.01032334379851818, + 0.011645245365798473, + 0.008213818073272705, + 0.009753232821822166, + 0.009661003947257996, + 0.0071225035935640335, + 0.009697163477540016, + 0.0038749182131141424, + 0.0045616053976118565, + 0.0029317864682525396, + 0.004375615157186985, + 0.0050065224058926105, + 0.0037848290521651506, + 0.008868709206581116, + 0.004489867482334375, + 0.007204522844403982, + 0.006018811371177435, + 0.007991071790456772, + 0.0064284345135092735, + 0.00848168134689331, + 0.005609965417534113, + 0.005652356427162886, + 0.006884180475026369, + 0.011877540498971939, + 0.0059931930154562, + 0.005590835586190224, + 0.005729279480874538, + 0.005231672897934914, + 0.0034913073759526014, + 0.007384322118014097, + 0.013389317318797112, + 0.017179349437355995, + 0.007146004121750593, + 0.007611789274960756, + 0.008002577349543571, + 0.0056712329387664795, + 0.007956080138683319, + 0.011981931515038013, + 0.015797778964042664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038089267909526825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1620182991027832, + 0.0809503048658371, + 0.1329522281885147, + 0.069397933781147, + 0.2059539407491684, + 0.0711006373167038, + 0.030487140640616417, + 0.08724918216466904, + 0.10360243916511536, + 0.05628791078925133 + ], + "credit_map": [ + 0.030382435768842697, + 0.015025284141302109, + 0.024143781512975693, + 0.019902030006051064, + 0.01965779811143875, + 0.013187387958168983, + 0.019910354167222977, + 0.01980922743678093, + 0.016067514196038246, + 0.008304200135171413, + 0.009378339163959026, + 0.007246450986713171, + 0.00897940993309021, + 0.013479892164468765, + 0.006718497257679701, + 0.010776003822684288, + 0.019380826503038406, + 0.015018334612250328, + 0.01811721734702587, + 0.0193051565438509, + 0.015653567388653755, + 0.014067157171666622, + 0.011515277437865734, + 0.019894687458872795, + 0.010320084169507027, + 0.00645082863047719, + 0.00810513086616993, + 0.008159289136528969, + 0.006696836091578007, + 0.007436209358274937, + 0.011329936794936657, + 0.010899617336690426, + 0.026593118906021118, + 0.029185213148593903, + 0.0269379410892725, + 0.03639281541109085, + 0.02235816791653633, + 0.02076600305736065, + 0.019837941974401474, + 0.023882737383246422, + 0.00994502566754818, + 0.01048960443586111, + 0.006643563974648714, + 0.009823937900364399, + 0.008615458384156227, + 0.008805591613054276, + 0.006324025336652994, + 0.010453426279127598, + 0.0038156204391270876, + 0.00534854969009757, + 0.0029498422518372536, + 0.0036470091436058283, + 0.004465878941118717, + 0.003344047348946333, + 0.0038820491172373295, + 0.003034142544493079, + 0.013532916083931923, + 0.00905931182205677, + 0.014806061051785946, + 0.011484571732580662, + 0.012496602721512318, + 0.008353160694241524, + 0.008685589767992496, + 0.008830973878502846, + 0.01568416692316532, + 0.008284704759716988, + 0.01574244536459446, + 0.011374829337000847, + 0.011082851327955723, + 0.010456465184688568, + 0.012829054147005081, + 0.018147919327020645, + 0.0071669383905828, + 0.005359035916626453, + 0.010032938793301582, + 0.006502367090433836, + 0.006780723575502634, + 0.006916957441717386, + 0.0063773770816624165, + 0.007151575293391943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0036149173974990845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot at eye level, captures the intimate moment of reunion in a sun-drenched kitchen. An older woman with soft, silver-white hair pulled back in a loose bun and a young man in his twenties recline at a rustic wooden table, embraced in a warm, gentle embrace. The woman, wearing a simple cream-colored blouse and a faded floral apron, has her eyes closed, a serene smile on her face, her wrinkled hands resting tenderly on the man's shoulders. The man, wearing a navy blue Henley shirt, wraps his arms firmly around her back, his face resting lightly against her temples, a look of relief and affection. The centerpiece of the scene is a freshly baked pie on a white ceramic plate, nestled directly between them; its golden-brown lattice crust gleams, a delicate plume of steam rising from its center. Golden afternoon sunlight pours through a nearby window, illuminating the scene, capturing the dancing dust particles in the air and casting soft, warm highlights on their faces and the polished table. The backdrop showcases a cozy, live-in kitchen with white cabinets and tile countertops, creating an atmosphere of deep tranquility, comfort, and heartfelt joy.", + "seed": 9811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_5824_seed_9811_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028751813620328904, + "vlm_logprob": -0.028751813620328904, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.028751813620328904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07296223193407059 + ], + "opposite_logprob_tokens": [ + -2.9479622840881348 + ], + "logprob_reward": -0.07296223193407059, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07296223193407059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3339931070804596, + 0.10329729318618774, + 0.052663568407297134, + 0.08985607326030731, + 0.07606309652328491, + 0.04964118078351021, + 0.09195975959300995, + 0.049997907131910324, + 0.11469726264476776, + 0.03783075511455536 + ], + "credit_map": [ + 0.05533597618341446, + 0.04208645224571228, + 0.03309423476457596, + 0.04695957899093628, + 0.044514432549476624, + 0.05052462965250015, + 0.030824851244688034, + 0.03065294399857521, + 0.013514506630599499, + 0.009469866752624512, + 0.010376173071563244, + 0.016736190766096115, + 0.0165973249822855, + 0.019231855869293213, + 0.008173731155693531, + 0.009197648614645004, + 0.00602909829467535, + 0.008727621287107468, + 0.005022202618420124, + 0.008314539678394794, + 0.00605805916711688, + 0.007066287100315094, + 0.004613716155290604, + 0.006832053419202566, + 0.013855217024683952, + 0.009949511848390102, + 0.009467840194702148, + 0.012346685864031315, + 0.017064442858099937, + 0.01095912978053093, + 0.008077651262283325, + 0.008135596290230751, + 0.009845084510743618, + 0.010155715048313141, + 0.008039958775043488, + 0.01454863604158163, + 0.00714719807729125, + 0.013057067058980465, + 0.006053052376955748, + 0.007216396275907755, + 0.005876423325389624, + 0.005472264718264341, + 0.004885911010205746, + 0.010628779418766499, + 0.005784541368484497, + 0.007228321395814419, + 0.004676078911870718, + 0.005088867619633675, + 0.014001002535223961, + 0.009678548201918602, + 0.011267201043665409, + 0.009296152740716934, + 0.014106723479926586, + 0.013369938358664513, + 0.009269678965210915, + 0.010970516130328178, + 0.004611216019839048, + 0.003917059861123562, + 0.006075833458453417, + 0.0075523993000388145, + 0.00691760191693902, + 0.005552193149924278, + 0.010406709276139736, + 0.004964897409081459, + 0.021136540919542313, + 0.011563758365809917, + 0.013477981090545654, + 0.010464468970894814, + 0.015474759042263031, + 0.01875193975865841, + 0.014035044237971306, + 0.009792779572308064, + 0.005773139651864767, + 0.0053253332152962685, + 0.004172833636403084, + 0.0050009191036224365, + 0.0065382253378629684, + 0.0039330837316811085, + 0.0033779998775571585, + 0.0037092245183885098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07296223193407059 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214574359357357 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214574359357357, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214574359357357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39692971110343933, + 0.08364664018154144, + 0.12890930473804474, + 0.03541257977485657, + 0.0793427899479866, + 0.0448848158121109, + 0.035183753818273544, + 0.03791642561554909, + 0.11671699583530426, + 0.04105694964528084 + ], + "credit_map": [ + 0.04701878875494003, + 0.03653845936059952, + 0.04171709343791008, + 0.06331325322389603, + 0.05683394521474838, + 0.06737106293439865, + 0.039350781589746475, + 0.04478634148836136, + 0.009475643746554852, + 0.007222683634608984, + 0.008297444321215153, + 0.014562379568815231, + 0.012684260495007038, + 0.015040633268654346, + 0.008074399083852768, + 0.008289202116429806, + 0.01917082630097866, + 0.01741623878479004, + 0.013537346385419369, + 0.016306649893522263, + 0.013108955696225166, + 0.014685011468827724, + 0.01927558332681656, + 0.015408706851303577, + 0.00320167257450521, + 0.0034079295583069324, + 0.0037699597887694836, + 0.005308463238179684, + 0.006084682419896126, + 0.005119128152728081, + 0.0036821099929511547, + 0.004838635679334402, + 0.011123462580144405, + 0.007017084397375584, + 0.009185736067593098, + 0.012091031298041344, + 0.006262240931391716, + 0.015435473062098026, + 0.008212661370635033, + 0.01001510675996542, + 0.005440308712422848, + 0.00892493687570095, + 0.004708093591034412, + 0.007053733803331852, + 0.004470237996429205, + 0.005201032850891352, + 0.003869102569296956, + 0.005217366851866245, + 0.006370231043547392, + 0.004172048531472683, + 0.004126159939914942, + 0.003531971713528037, + 0.005504105240106583, + 0.00440194271504879, + 0.0025679864920675755, + 0.0045093088410794735, + 0.004272988066077232, + 0.0027885360177606344, + 0.004749833606183529, + 0.006863264832645655, + 0.003609120612964034, + 0.0075406888499855995, + 0.0038144986610859632, + 0.004277493804693222, + 0.019536012783646584, + 0.01330418698489666, + 0.014732200652360916, + 0.013297379948198795, + 0.018754614517092705, + 0.018897444009780884, + 0.010859244503080845, + 0.007335907779633999, + 0.007766357623040676, + 0.00619553541764617, + 0.004622985143214464, + 0.0056252735666930676, + 0.0051638404838740826, + 0.004709430038928986, + 0.0035196770913898945, + 0.0034538530744612217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.014214574359357357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3723244071006775, + 0.20323103666305542, + 0.08318882435560226, + 0.044217418879270554, + 0.04791611060500145, + 0.0484125092625618, + 0.03823992237448692, + 0.042271051555871964, + 0.05326816812157631, + 0.06693051755428314 + ], + "credit_map": [ + 0.053735602647066116, + 0.03816822171211243, + 0.03626801073551178, + 0.04011145606637001, + 0.0734938383102417, + 0.043855734169483185, + 0.03372829407453537, + 0.0529632531106472, + 0.02517417073249817, + 0.018726758658885956, + 0.02302243374288082, + 0.03190279006958008, + 0.02854476124048233, + 0.034271009266376495, + 0.015593387186527252, + 0.025995733216404915, + 0.00814980361610651, + 0.012517408467829227, + 0.008569353260099888, + 0.012840017676353455, + 0.008616224862635136, + 0.009628253988921642, + 0.008599529042840004, + 0.014268235303461552, + 0.0046200258657336235, + 0.004928241949528456, + 0.004943469539284706, + 0.005441735964268446, + 0.00476386584341526, + 0.007234879769384861, + 0.00533608952537179, + 0.0069491080939769745, + 0.004702345468103886, + 0.004410846158862114, + 0.005013012792915106, + 0.008095120079815388, + 0.007542707491666079, + 0.007932905107736588, + 0.004737230949103832, + 0.005481944419443607, + 0.009705631993710995, + 0.004579885397106409, + 0.0064374180510640144, + 0.007404224947094917, + 0.006376722827553749, + 0.0068502649664878845, + 0.0032374036964029074, + 0.0038209580816328526, + 0.006780214142054319, + 0.004584018141031265, + 0.003849254921078682, + 0.004841177724301815, + 0.005603404249995947, + 0.005188456736505032, + 0.0031225939746946096, + 0.004270805045962334, + 0.006793507374823093, + 0.0032779069151729345, + 0.0060949781909585, + 0.005678392015397549, + 0.004358259961009026, + 0.007643301971256733, + 0.005162299145013094, + 0.003262406215071678, + 0.007179367821663618, + 0.005845210049301386, + 0.006356682162731886, + 0.006822494789958, + 0.01077347993850708, + 0.007849667221307755, + 0.004455620422959328, + 0.003985644783824682, + 0.01096239686012268, + 0.006916931830346584, + 0.006602380890399218, + 0.012409533374011517, + 0.005777124315500259, + 0.010369693860411644, + 0.008243687450885773, + 0.005648771300911903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012536883354187012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809810429811478 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.03809810429811478, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809810429811478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3195022642612457, + 0.16481010615825653, + 0.05421077460050583, + 0.05091268941760063, + 0.1704634428024292, + 0.04408938065171242, + 0.050429679453372955, + 0.08770871162414551, + 0.028886059299111366, + 0.028986848890781403 + ], + "credit_map": [ + 0.05086637660861015, + 0.02309034951031208, + 0.03553677350282669, + 0.04488491639494896, + 0.04421735182404518, + 0.06577413529157639, + 0.029713520780205727, + 0.025418857112526894, + 0.023741869255900383, + 0.011349152773618698, + 0.02124379575252533, + 0.033064670860767365, + 0.011966669000685215, + 0.03749699890613556, + 0.012636483646929264, + 0.013310479931533337, + 0.010780232958495617, + 0.00518777035176754, + 0.005521413404494524, + 0.00863981805741787, + 0.004509063437581062, + 0.008372385054826736, + 0.004085555672645569, + 0.00711454451084137, + 0.009593995288014412, + 0.00363673223182559, + 0.005627461243420839, + 0.00788476224988699, + 0.00543075380846858, + 0.007807083427906036, + 0.005557254422456026, + 0.005374651402235031, + 0.030568057671189308, + 0.010585934855043888, + 0.022496376186609268, + 0.026846950873732567, + 0.011991572566330433, + 0.03837191313505173, + 0.012844134122133255, + 0.01675853133201599, + 0.004981957841664553, + 0.004095477517694235, + 0.004997551441192627, + 0.008027291856706142, + 0.004140523262321949, + 0.009610319510102272, + 0.003584182122722268, + 0.004652082454413176, + 0.008326588198542595, + 0.005471132695674896, + 0.006512784864753485, + 0.004971358925104141, + 0.008082308806478977, + 0.00931152980774641, + 0.0038089528679847717, + 0.003945028409361839, + 0.01502592395991087, + 0.0061092134565114975, + 0.01104042585939169, + 0.010364990681409836, + 0.00690243998542428, + 0.02279357984662056, + 0.008128711022436619, + 0.007343432866036892, + 0.003141452092677355, + 0.0038507350254803896, + 0.0036789702717214823, + 0.0033501561265438795, + 0.005292889196425676, + 0.004081378225237131, + 0.002809870056807995, + 0.0026806097012013197, + 0.004480164498090744, + 0.004514062311500311, + 0.003328247694298625, + 0.004515449050813913, + 0.004102629609405994, + 0.0031383195891976357, + 0.002800856251269579, + 0.0021071212831884623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03809810429811478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2159719318151474, + 0.08383029699325562, + 0.07162556052207947, + 0.07077907025814056, + 0.09339514374732971, + 0.17772191762924194, + 0.10116320848464966, + 0.04856432229280472, + 0.07019160687923431, + 0.06675689667463303 + ], + "credit_map": [ + 0.03538885712623596, + 0.022363651543855667, + 0.021847449243068695, + 0.0360073484480381, + 0.025813428685069084, + 0.035255543887615204, + 0.015908289700746536, + 0.023387357592582703, + 0.011664297431707382, + 0.006678808946162462, + 0.010609395802021027, + 0.010795275680720806, + 0.010123813524842262, + 0.017074907198548317, + 0.00922095961868763, + 0.00766283692792058, + 0.013363698497414589, + 0.006716515403240919, + 0.007810382638126612, + 0.010028126649558544, + 0.008536717854440212, + 0.009822975844144821, + 0.007456022314727306, + 0.007891112938523293, + 0.011110437102615833, + 0.006056572310626507, + 0.007812823168933392, + 0.01022329181432724, + 0.010884851217269897, + 0.011277697049081326, + 0.006252451799809933, + 0.0071609485894441605, + 0.01661546900868416, + 0.014052909798920155, + 0.00904989056289196, + 0.011378096416592598, + 0.01332018617540598, + 0.0136699378490448, + 0.006403645966202021, + 0.008905005641281605, + 0.032903771847486496, + 0.019392013549804688, + 0.018153075128793716, + 0.024766063317656517, + 0.02133725769817829, + 0.0287004541605711, + 0.014956895262002945, + 0.017512381076812744, + 0.012577743269503117, + 0.01208746898919344, + 0.011625608429312706, + 0.013559619896113873, + 0.01747680827975273, + 0.010969379916787148, + 0.009532488882541656, + 0.013334091752767563, + 0.00808318704366684, + 0.004638053011149168, + 0.005982796661555767, + 0.005099078174680471, + 0.00844824593514204, + 0.005051329731941223, + 0.0070300018414855, + 0.004231628496199846, + 0.010923919267952442, + 0.008190851658582687, + 0.007810801267623901, + 0.009258029982447624, + 0.013292760588228703, + 0.008793472312390804, + 0.005375996232032776, + 0.006545775104314089, + 0.012461026199162006, + 0.006148333661258221, + 0.007880192250013351, + 0.01000146847218275, + 0.009434729814529419, + 0.00896733719855547, + 0.006128119304776192, + 0.00573568744584918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0059472741559147835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, shot at eye level, captures the intimate moment of reunion in a sun-drenched kitchen. An older woman with soft, silver-white hair pulled back in a loose bun and a young man in his twenties recline at a rustic wooden table, embraced in a warm, gentle embrace. The woman, wearing a simple cream-colored blouse and a faded floral apron, has her eyes closed, a serene smile on her face, her wrinkled hands resting tenderly on the man's shoulders. The man, wearing a navy blue Henley shirt, wraps his arms firmly around her back, his face resting lightly against her temples, a look of relief and affection. The centerpiece of the scene is a freshly baked pie on a white ceramic plate, nestled directly between them; its golden-brown lattice crust gleams, a delicate plume of steam rising from its center. Golden afternoon sunlight pours through a nearby window, illuminating the scene, capturing the dancing dust particles in the air and casting soft, warm highlights on their faces and the polished table. The backdrop showcases a cozy, live-in kitchen with white cabinets and tile countertops, creating an atmosphere of deep tranquility, comfort, and heartfelt joy.", + "seed": 9812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_11/index_5824_seed_9812_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027405041828751565, + "vlm_logprob": -0.027405041828751565, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.027405041828751565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06908317655324936 + ], + "opposite_logprob_tokens": [ + -3.6940832138061523 + ], + "logprob_reward": -0.06908317655324936, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.06908317655324936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.143377885222435, + 0.06468729674816132, + 0.03794337436556816, + 0.03946201130747795, + 0.03875523433089256, + 0.05433638021349907, + 0.08525306731462479, + 0.03277738392353058, + 0.04897136241197586, + 0.4544360041618347 + ], + "credit_map": [ + 0.015243140049278736, + 0.020635074004530907, + 0.012668232433497906, + 0.02390873059630394, + 0.016049066558480263, + 0.01783907227218151, + 0.020152201876044273, + 0.01688236929476261, + 0.007562630344182253, + 0.008664451539516449, + 0.005924114026129246, + 0.008443075232207775, + 0.008430812507867813, + 0.009616139344871044, + 0.007998061366379261, + 0.008048009127378464, + 0.0034461822360754013, + 0.005360852926969528, + 0.0037616139743477106, + 0.006205628626048565, + 0.004375053104013205, + 0.004946285858750343, + 0.005619970615953207, + 0.004227784927934408, + 0.004468744620680809, + 0.005542114842683077, + 0.0034195431508123875, + 0.006230519153177738, + 0.004482317715883255, + 0.004785866476595402, + 0.00651325611397624, + 0.004019647371023893, + 0.0038942701648920774, + 0.006091377232223749, + 0.0030772360041737556, + 0.005496359895914793, + 0.006359871011227369, + 0.004833640530705452, + 0.004903586581349373, + 0.004098894074559212, + 0.005566999316215515, + 0.007504126522690058, + 0.005515768192708492, + 0.011897899210453033, + 0.006096167489886284, + 0.00545261288061738, + 0.007272479124367237, + 0.00503033259883523, + 0.009413409046828747, + 0.012248676270246506, + 0.008090279065072536, + 0.014040925540030003, + 0.008647309616208076, + 0.01133654359728098, + 0.012415670789778233, + 0.009060252457857132, + 0.00318116694688797, + 0.0034253753256052732, + 0.0030560861341655254, + 0.005449433345347643, + 0.0055290707387030125, + 0.004519288428127766, + 0.003965158015489578, + 0.0036518052220344543, + 0.004213583189994097, + 0.005457868799567223, + 0.004286668729037046, + 0.0068952348083257675, + 0.007839410565793514, + 0.005500256549566984, + 0.004269327037036419, + 0.010509009473025799, + 0.019068574532866478, + 0.03270748630166054, + 0.015376441180706024, + 0.020364386960864067, + 0.15658633410930634, + 0.04090224951505661, + 0.03422551229596138, + 0.13520501554012299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.06908317655324936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659903891384602 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.008659903891384602, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659903891384602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18223489820957184, + 0.08597766607999802, + 0.04033634811639786, + 0.017932618036866188, + 0.02162080630660057, + 0.03461700305342674, + 0.027627678588032722, + 0.01888112910091877, + 0.017508722841739655, + 0.5532631278038025 + ], + "credit_map": [ + 0.013837765902280807, + 0.02386372908949852, + 0.012368296273052692, + 0.029632842168211937, + 0.040177442133426666, + 0.017279142513871193, + 0.022455500438809395, + 0.02262016199529171, + 0.007168476469814777, + 0.013132118619978428, + 0.007529997732490301, + 0.014965515583753586, + 0.012689963914453983, + 0.010564508847892284, + 0.010182792320847511, + 0.00974428839981556, + 0.003030553925782442, + 0.005460699088871479, + 0.004023985005915165, + 0.004451692570000887, + 0.004595745354890823, + 0.007505021523684263, + 0.00490892818197608, + 0.0063597215339541435, + 0.002015112666413188, + 0.00223531574010849, + 0.001560068572871387, + 0.0026513012126088142, + 0.002921333070844412, + 0.0017352435970678926, + 0.0028050101827830076, + 0.002009232062846422, + 0.0020904592238366604, + 0.0031076367013156414, + 0.002355285221710801, + 0.003086263546720147, + 0.0030706494580954313, + 0.002426864579319954, + 0.0028265086002647877, + 0.0026571399066597223, + 0.002522609895095229, + 0.004126403480768204, + 0.0035628199111670256, + 0.004759086761623621, + 0.005622556433081627, + 0.0036487148609012365, + 0.004881566856056452, + 0.005493244156241417, + 0.0028354860842227936, + 0.003903853939846158, + 0.002696154173463583, + 0.003955058753490448, + 0.003220504382625222, + 0.0036729013081640005, + 0.003935488872230053, + 0.003408229211345315, + 0.0013189006131142378, + 0.002161554293707013, + 0.002238225657492876, + 0.0031504896469414234, + 0.0031328825280070305, + 0.0017084416467696428, + 0.00209630886092782, + 0.0030743235256522894, + 0.0014242014149203897, + 0.002360524842515588, + 0.0014649257063865662, + 0.003670755075290799, + 0.0030062224250286818, + 0.0016374604310840368, + 0.002143541583791375, + 0.0018010896164923906, + 0.023012898862361908, + 0.0380672812461853, + 0.01818985491991043, + 0.022147325798869133, + 0.19499965012073517, + 0.051710959523916245, + 0.041607920080423355, + 0.1635272055864334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008659903891384602 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538767419755459 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012538767419755459, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538767419755459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15268029272556305, + 0.10167752951383591, + 0.06993497908115387, + 0.0461227223277092, + 0.038107842206954956, + 0.04039417952299118, + 0.04749341681599617, + 0.04914778098464012, + 0.04641615226864815, + 0.4080251455307007 + ], + "credit_map": [ + 0.014136373065412045, + 0.018478117883205414, + 0.015953246504068375, + 0.01694163866341114, + 0.022250907495617867, + 0.027548937126994133, + 0.018517544493079185, + 0.018853534013032913, + 0.009114377200603485, + 0.013384510762989521, + 0.009608326479792595, + 0.01251731812953949, + 0.012003671377897263, + 0.019943999126553535, + 0.012466518208384514, + 0.012638808228075504, + 0.0064580803737044334, + 0.006594543810933828, + 0.005586378276348114, + 0.006848073098808527, + 0.007849703542888165, + 0.01845443993806839, + 0.006988910026848316, + 0.011154844425618649, + 0.004385942593216896, + 0.005680053494870663, + 0.004505075514316559, + 0.006579599343240261, + 0.006480934098362923, + 0.006630783900618553, + 0.006994333583861589, + 0.004866002593189478, + 0.003580603050068021, + 0.003545367391780019, + 0.00307503342628479, + 0.004704877268522978, + 0.004136649426072836, + 0.008601529523730278, + 0.00419424707069993, + 0.006269535515457392, + 0.004097464960068464, + 0.004358219914138317, + 0.0032396800816059113, + 0.006842335686087608, + 0.004399606492370367, + 0.00680819246917963, + 0.005458432249724865, + 0.0051902481354773045, + 0.0043913936242461205, + 0.007003806997090578, + 0.004928999580442905, + 0.006382005754858255, + 0.004474710673093796, + 0.006914385594427586, + 0.00719232065603137, + 0.006205793935805559, + 0.003631969913840294, + 0.006754330359399319, + 0.004083449020981789, + 0.007708792109042406, + 0.007394307758659124, + 0.0063463663682341576, + 0.008220305666327477, + 0.00500826071947813, + 0.0037962866481393576, + 0.005587903782725334, + 0.004102244973182678, + 0.007776358164846897, + 0.007550524082034826, + 0.005027284845709801, + 0.00763340201228857, + 0.0049421461299061775, + 0.016138719394803047, + 0.028230788186192513, + 0.012847894802689552, + 0.015270196832716465, + 0.14538739621639252, + 0.039412859827280045, + 0.031217467039823532, + 0.11951980739831924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012538767419755459 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381057932972908 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.0381057932972908, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381057932972908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631452977657318, + 0.19556614756584167, + 0.16574712097644806, + 0.02019839733839035, + 0.05726511403918266, + 0.10533369332551956, + 0.04670669138431549, + 0.05515563115477562, + 0.02123093232512474, + 0.06965091824531555 + ], + "credit_map": [ + 0.030956637114286423, + 0.03337177261710167, + 0.02343256026506424, + 0.026870086789131165, + 0.034324754029512405, + 0.05155409500002861, + 0.03623615577816963, + 0.026399249210953712, + 0.01602671667933464, + 0.020626302808523178, + 0.014500414952635765, + 0.028742387890815735, + 0.027527479454874992, + 0.03450988605618477, + 0.03135437145829201, + 0.022278595715761185, + 0.014851471409201622, + 0.017055563628673553, + 0.013441148214042187, + 0.022783188149333, + 0.019923994317650795, + 0.03683675825595856, + 0.0231004785746336, + 0.017754511907696724, + 0.0015618869801983237, + 0.002281119115650654, + 0.0014297987800091505, + 0.002710344735532999, + 0.0028712861239910126, + 0.0030727400444447994, + 0.003487776033580303, + 0.002783446107059717, + 0.004678066354244947, + 0.005610598251223564, + 0.005210652481764555, + 0.007257341872900724, + 0.007676129229366779, + 0.011423954740166664, + 0.008172604255378246, + 0.007235766388475895, + 0.010032477788627148, + 0.01032518595457077, + 0.006718362215906382, + 0.011857678182423115, + 0.012440130114555359, + 0.02608942613005638, + 0.014788363128900528, + 0.013082072138786316, + 0.0038998350501060486, + 0.005531967151910067, + 0.00450611999258399, + 0.006394209340214729, + 0.0050639077089726925, + 0.008980202488601208, + 0.00567011721432209, + 0.006660329643636942, + 0.004288212396204472, + 0.006912367884069681, + 0.004248024895787239, + 0.00740407221019268, + 0.008772220462560654, + 0.008590731769800186, + 0.00960660632699728, + 0.005333395209163427, + 0.0018159213941544294, + 0.0021656688768416643, + 0.0014757422031834722, + 0.0028569542337208986, + 0.003500948427245021, + 0.0036720612552016973, + 0.0032691452652215958, + 0.0024744898546487093, + 0.004783868324011564, + 0.005977041088044643, + 0.0031654336489737034, + 0.005866631865501404, + 0.017165010794997215, + 0.013774195685982704, + 0.007961477153003216, + 0.010957262478768826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0381057932972908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are an older woman with silver-white hair in a bun and a young man in a navy blue Henley shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a freshly baked pie with a golden-brown lattice crust on a white ceramic plate positioned between the two people?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched kitchen with white cabinets and tile countertops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is golden afternoon sunlight visible pouring through a window, illuminating dust particles and casting highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637567982077599 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637567982077599, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637567982077599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2997867166996002, + 0.1098618432879448, + 0.06915697455406189, + 0.04048272967338562, + 0.0512106791138649, + 0.06399265676736832, + 0.08909709006547928, + 0.05923973023891449, + 0.0468563586473465, + 0.1703152060508728 + ], + "credit_map": [ + 0.026754649356007576, + 0.03728323057293892, + 0.030713027343153954, + 0.03431834280490875, + 0.03914305195212364, + 0.05173845216631889, + 0.045277226716279984, + 0.0345587357878685, + 0.010918381623923779, + 0.012577281333506107, + 0.009328719228506088, + 0.013659896329045296, + 0.013822968117892742, + 0.01523836050182581, + 0.019276266917586327, + 0.015039966441690922, + 0.006434279959648848, + 0.008080326952040195, + 0.0060940817929804325, + 0.007516176905483007, + 0.007190004922449589, + 0.01581585966050625, + 0.009146492928266525, + 0.008879752829670906, + 0.0038083645049482584, + 0.004567510914057493, + 0.00342164421454072, + 0.004829053767025471, + 0.00488381739705801, + 0.0060840570367872715, + 0.008069822564721107, + 0.004818461835384369, + 0.004848749376833439, + 0.005991131532937288, + 0.00448176683858037, + 0.005787370260804892, + 0.006614587269723415, + 0.0067423549480736256, + 0.010310261510312557, + 0.006434459239244461, + 0.005911323241889477, + 0.007514203432947397, + 0.00555682135745883, + 0.008494267240166664, + 0.009375803172588348, + 0.010344165377318859, + 0.009286834858357906, + 0.007509244605898857, + 0.009265935979783535, + 0.011383873410522938, + 0.0084909126162529, + 0.014569699764251709, + 0.01131491456180811, + 0.013986555859446526, + 0.010426191613078117, + 0.009659016504883766, + 0.004414839204400778, + 0.007164269685745239, + 0.004253793507814407, + 0.007817454636096954, + 0.011248933151364326, + 0.00593150220811367, + 0.009649376384913921, + 0.008759560994803905, + 0.00386875425465405, + 0.0058641983196139336, + 0.00331235583871603, + 0.006886562332510948, + 0.007467432878911495, + 0.0038981891702860594, + 0.007069423794746399, + 0.008489442057907581, + 0.008031441830098629, + 0.01191800832748413, + 0.006776344031095505, + 0.00944066047668457, + 0.056378260254859924, + 0.01698216423392296, + 0.014040132984519005, + 0.04674818739295006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008637567982077599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_tensor([6670, 6670])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_tensor([6670, 6670])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..097da6182b45d3879e1fa38d91ee431c742997f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_tensor([6670, 6670])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot, using a bullet-time effect, captures in astonishing detail the moment an arrow completely pierces a red apple. The arrow, with a sleek black carbon fiber shaft and a sharp, gleaming steel tip, has just burst from the other side of the apple. The perfectly ripe apple, its skin a vibrant crimson, is now shattered inward by the immense force of the impact. A spectacular, three-dimensional corona of juice and pulp erupts around the arrow, with thousands of individual, crystalline droplets and tiny fibers of flesh frozen in mid-air. Some larger droplets remain perfectly spherical due to surface tension, while others are stretched into delicate threads of liquid. The entire scene is set against a pure black, non-reflective background. Sharp, bright, high-speed studio lighting illuminates every tiny detail from the side, capturing every facet of the splashing droplets, making them sparkle like countless tiny diamonds while creating crisp highlights on the smooth skin of the apple and the metal surface of the arrow. The entire picture presents an extreme surreal texture, freezing a moment of destruction full of violent aesthetics with scientific precision and artistic clarity, and the details are extremely realistic.", + "seed": 15306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_6670_seed_15306_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031417943979613484, + "vlm_logprob": -0.031417943979613484, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.031417943979613484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023110216483473778 + ], + "opposite_logprob_tokens": [ + -3.8981101512908936 + ], + "logprob_reward": -0.023110216483473778, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023110216483473778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12335219979286194, + 0.07169042527675629, + 0.14351941645145416, + 0.135601207613945, + 0.03462379798293114, + 0.11746731400489807, + 0.23089417815208435, + 0.08173026889562607, + 0.03547833487391472, + 0.025642860680818558 + ], + "credit_map": [ + 0.01694398559629917, + 0.008971610106527805, + 0.017716243863105774, + 0.010512233711779118, + 0.012490727938711643, + 0.027403350919485092, + 0.013570299372076988, + 0.01574375107884407, + 0.011291926726698875, + 0.004850895144045353, + 0.012407130561769009, + 0.006498829927295446, + 0.006554301828145981, + 0.014163847081363201, + 0.007963765412569046, + 0.007959726266562939, + 0.029819171875715256, + 0.012167590670287609, + 0.01930307224392891, + 0.011765685863792896, + 0.012549729086458683, + 0.03207375109195709, + 0.010775340721011162, + 0.015065065585076809, + 0.020632289350032806, + 0.011527911759912968, + 0.014198042452335358, + 0.017685001716017723, + 0.012113218195736408, + 0.02856150083243847, + 0.016279244795441628, + 0.014604008756577969, + 0.004237615503370762, + 0.006258974317461252, + 0.006272902246564627, + 0.003296585287898779, + 0.0047234841622412205, + 0.004307519178837538, + 0.0025650185998529196, + 0.002961698453873396, + 0.034899380058050156, + 0.007736894302070141, + 0.030057232826948166, + 0.008183294907212257, + 0.007181879132986069, + 0.011320543475449085, + 0.009042066521942616, + 0.009046021848917007, + 0.029580196365714073, + 0.02750309556722641, + 0.037924349308013916, + 0.03256251662969589, + 0.024691615253686905, + 0.03746702894568443, + 0.022672411054372787, + 0.01849295198917389, + 0.010750285349786282, + 0.007575174327939749, + 0.009626518003642559, + 0.012284986674785614, + 0.008614971302449703, + 0.012007315643131733, + 0.011672917753458023, + 0.009198104962706566, + 0.0066030011512339115, + 0.00201540719717741, + 0.004694568924605846, + 0.0031002345494925976, + 0.002700702752918005, + 0.0057458034716546535, + 0.004986482672393322, + 0.005632133688777685, + 0.005063570570200682, + 0.0024321828968822956, + 0.004165072925388813, + 0.0027937719132751226, + 0.0021258345805108547, + 0.0036581819877028465, + 0.0030924503225833178, + 0.0023117968812584877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.023110216483473778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871518537402153 + ], + "opposite_logprob_tokens": [ + -3.048715114593506 + ], + "logprob_reward": -0.04871518537402153, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871518537402153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1251956820487976, + 0.17298178374767303, + 0.06352882087230682, + 0.052928850054740906, + 0.09150132536888123, + 0.16513299942016602, + 0.12078647315502167, + 0.1341879963874817, + 0.041137054562568665, + 0.0326189324259758 + ], + "credit_map": [ + 0.017658965662121773, + 0.009928110055625439, + 0.0195936132222414, + 0.015443638898432255, + 0.007360674440860748, + 0.02074679732322693, + 0.013192019425332546, + 0.021271873265504837, + 0.03154461085796356, + 0.011357889510691166, + 0.024127868935465813, + 0.019583966583013535, + 0.017930768430233, + 0.024218179285526276, + 0.020766951143741608, + 0.023451542481780052, + 0.009758423082530499, + 0.005897102877497673, + 0.008401177823543549, + 0.007463363930583, + 0.005214761942625046, + 0.013793849386274815, + 0.005670658312737942, + 0.0073294807225465775, + 0.00807962752878666, + 0.004989840090274811, + 0.00679024076089263, + 0.007761666085571051, + 0.004275425802916288, + 0.008191407658159733, + 0.004925675690174103, + 0.007914965972304344, + 0.01351328007876873, + 0.01462840847671032, + 0.01466541737318039, + 0.010134254582226276, + 0.008605258539319038, + 0.013145825825631618, + 0.008592437021434307, + 0.00821644440293312, + 0.03612662851810455, + 0.008249854668974876, + 0.05147086828947067, + 0.010135742835700512, + 0.007002323865890503, + 0.024039020761847496, + 0.010233388282358646, + 0.01787516660988331, + 0.01567741669714451, + 0.012424458749592304, + 0.020191777497529984, + 0.015467112883925438, + 0.011621100828051567, + 0.023360472172498703, + 0.011290543712675571, + 0.010753590613603592, + 0.025901300832629204, + 0.009386170655488968, + 0.0202201958745718, + 0.01607019267976284, + 0.01051655225455761, + 0.01916692405939102, + 0.016472477465867996, + 0.01645417883992195, + 0.01023358479142189, + 0.0025352174416184425, + 0.00653415359556675, + 0.003949309233576059, + 0.0026483817491680384, + 0.006385594606399536, + 0.0041694496758282185, + 0.004681362304836512, + 0.0064224074594676495, + 0.002631762530654669, + 0.004698276519775391, + 0.004845047369599342, + 0.002127481624484062, + 0.004459756426513195, + 0.0035870575811713934, + 0.0038471431471407413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.04871518537402153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199178665876389 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199178665876389, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199178665876389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06572690606117249, + 0.0662192702293396, + 0.1856328696012497, + 0.25803545117378235, + 0.06051251292228699, + 0.13611510396003723, + 0.13593079149723053, + 0.05277099832892418, + 0.023958778008818626, + 0.01509736105799675 + ], + "credit_map": [ + 0.009879549965262413, + 0.005082328803837299, + 0.010355448350310326, + 0.007916931994259357, + 0.005893290042877197, + 0.008861400187015533, + 0.008695877157151699, + 0.00904207956045866, + 0.011904148384928703, + 0.004127497784793377, + 0.011045985855162144, + 0.006761476397514343, + 0.006079146638512611, + 0.01237429492175579, + 0.006767435930669308, + 0.007159283384680748, + 0.036017242819070816, + 0.015341940335929394, + 0.02423950657248497, + 0.018506063148379326, + 0.016644585877656937, + 0.038971368223428726, + 0.015874169766902924, + 0.020037978887557983, + 0.03498096764087677, + 0.027282359078526497, + 0.02767784893512726, + 0.04062158986926079, + 0.022438064217567444, + 0.046582527458667755, + 0.03151965141296387, + 0.026932448148727417, + 0.009686693549156189, + 0.006451407913118601, + 0.011718298308551311, + 0.006398495752364397, + 0.004446780774742365, + 0.01012000534683466, + 0.005540455691516399, + 0.006150378379970789, + 0.04115436226129532, + 0.005818005185574293, + 0.04411854222416878, + 0.007715790066868067, + 0.005261244717985392, + 0.01199696958065033, + 0.00768924318253994, + 0.012360943481326103, + 0.01803663559257984, + 0.019037267193198204, + 0.023425906896591187, + 0.02161092683672905, + 0.01390004064887762, + 0.016683978959918022, + 0.013506067916750908, + 0.00972997210919857, + 0.008094022050499916, + 0.0055761015973985195, + 0.0063848551362752914, + 0.006915039848536253, + 0.005489489063620567, + 0.00901145301759243, + 0.006111009046435356, + 0.005189032759517431, + 0.004314956720918417, + 0.0018315422348678112, + 0.0030103460885584354, + 0.002074663760140538, + 0.0021180924959480762, + 0.004141425713896751, + 0.002445891499519348, + 0.0040218583308160305, + 0.0023892284370958805, + 0.0013956649927422404, + 0.0022198797669261694, + 0.002763188909739256, + 0.0013438139576464891, + 0.002086070366203785, + 0.0016182613326236606, + 0.0012812528293579817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.014199178665876389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001914829015732 + ], + "opposite_logprob_tokens": [ + -2.695019245147705 + ], + "logprob_reward": -0.07001914829015732, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001914829015732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03890536352992058, + 0.20165322721004486, + 0.0616789311170578, + 0.028031060472130775, + 0.1526060700416565, + 0.08981133252382278, + 0.23983342945575714, + 0.04128255695104599, + 0.03204507753252983, + 0.1141529455780983 + ], + "credit_map": [ + 0.006982572376728058, + 0.0031922406051307917, + 0.005384569056332111, + 0.0037786560133099556, + 0.004477259237319231, + 0.005720187909901142, + 0.005046230740845203, + 0.004323645494878292, + 0.03811941668391228, + 0.013013070449233055, + 0.025670073926448822, + 0.020807193592190742, + 0.021828152239322662, + 0.02782811038196087, + 0.03155891224741936, + 0.02282831072807312, + 0.010795269161462784, + 0.005852820817381144, + 0.008929102681577206, + 0.007078040391206741, + 0.005804257467389107, + 0.011034565977752209, + 0.00636700727045536, + 0.005817870609462261, + 0.004078466910868883, + 0.003454308258369565, + 0.003548676148056984, + 0.0038564151618629694, + 0.003512165741994977, + 0.003738563973456621, + 0.0026929015293717384, + 0.0031495646107941866, + 0.028353044763207436, + 0.012262532487511635, + 0.03612038120627403, + 0.01316195446997881, + 0.01185907144099474, + 0.024170061573386192, + 0.012402884662151337, + 0.014276145026087761, + 0.026993919163942337, + 0.0034458842128515244, + 0.03384983912110329, + 0.004767595324665308, + 0.0026699805166572332, + 0.006804023869335651, + 0.004013010300695896, + 0.0072670807130634785, + 0.0323074534535408, + 0.03079123981297016, + 0.04877127334475517, + 0.03148949146270752, + 0.02468973584473133, + 0.039513591676950455, + 0.018586644902825356, + 0.013684000819921494, + 0.007700059562921524, + 0.003050437429919839, + 0.005302160046994686, + 0.004855215083807707, + 0.0036898450925946236, + 0.004653929732739925, + 0.0060986243188381195, + 0.005932285916060209, + 0.0075704255141317844, + 0.0022372445091605186, + 0.0051444340497255325, + 0.0029298800509423018, + 0.0027031665667891502, + 0.004229257814586163, + 0.003125048941001296, + 0.004105620551854372, + 0.02112693153321743, + 0.0088507616892457, + 0.019444094970822334, + 0.014391680248081684, + 0.007898698560893536, + 0.023892659693956375, + 0.008295566774904728, + 0.010252554900944233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.07001914829015732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010459910845384002 + ], + "opposite_logprob_tokens": [ + -6.876046180725098 + ], + "logprob_reward": -0.0010459910845384002, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010459910845384002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08902750164270401, + 0.15151767432689667, + 0.11961537599563599, + 0.1277834177017212, + 0.1215764582157135, + 0.1033196970820427, + 0.07007545232772827, + 0.08118594437837601, + 0.04967019706964493, + 0.08622833341360092 + ], + "credit_map": [ + 0.01283726654946804, + 0.007232713047415018, + 0.011134006083011627, + 0.01309935748577118, + 0.00774335814639926, + 0.010719084180891514, + 0.012724589556455612, + 0.013537127524614334, + 0.025986984372138977, + 0.009927201084792614, + 0.02206580899655819, + 0.017456911504268646, + 0.01710602641105652, + 0.020787755027413368, + 0.024355432018637657, + 0.013831548392772675, + 0.022590195760130882, + 0.011635971255600452, + 0.016296975314617157, + 0.011222267523407936, + 0.011448809877038002, + 0.026871908456087112, + 0.008267851546406746, + 0.011281401850283146, + 0.026949217543005943, + 0.011624998413026333, + 0.014838269911706448, + 0.012102014385163784, + 0.010731237940490246, + 0.030465848743915558, + 0.00872059166431427, + 0.012351244688034058, + 0.021088561043143272, + 0.01025642454624176, + 0.02642427571117878, + 0.01256355457007885, + 0.009722455404698849, + 0.019005801528692245, + 0.010117737576365471, + 0.012397648766636848, + 0.027680573984980583, + 0.004907573573291302, + 0.031325627118349075, + 0.005859172437340021, + 0.006314476951956749, + 0.011690502054989338, + 0.007440901827067137, + 0.00810086540877819, + 0.012374219484627247, + 0.006654436234384775, + 0.020207200199365616, + 0.008177199400961399, + 0.0045111170038580894, + 0.0061028944328427315, + 0.007382582873106003, + 0.004665805026888847, + 0.012273636646568775, + 0.006292782258242369, + 0.011284253560006618, + 0.009526635520160198, + 0.00876585952937603, + 0.010956548154354095, + 0.011982438154518604, + 0.01010379008948803, + 0.00969116110354662, + 0.0030886761378496885, + 0.007935956120491028, + 0.007522594183683395, + 0.0030778301879763603, + 0.007679562550038099, + 0.004146389663219452, + 0.006528027355670929, + 0.020417558029294014, + 0.005520492326468229, + 0.014339156448841095, + 0.00898442231118679, + 0.006953798234462738, + 0.012980985455214977, + 0.00768505921587348, + 0.00934686604887247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.0010459910845384002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, using a bullet-time effect, captures in astonishing detail the moment an arrow completely pierces a red apple. The arrow, with a sleek black carbon fiber shaft and a sharp, gleaming steel tip, has just burst from the other side of the apple. The perfectly ripe apple, its skin a vibrant crimson, is now shattered inward by the immense force of the impact. A spectacular, three-dimensional corona of juice and pulp erupts around the arrow, with thousands of individual, crystalline droplets and tiny fibers of flesh frozen in mid-air. Some larger droplets remain perfectly spherical due to surface tension, while others are stretched into delicate threads of liquid. The entire scene is set against a pure black, non-reflective background. Sharp, bright, high-speed studio lighting illuminates every tiny detail from the side, capturing every facet of the splashing droplets, making them sparkle like countless tiny diamonds while creating crisp highlights on the smooth skin of the apple and the metal surface of the arrow. The entire picture presents an extreme surreal texture, freezing a moment of destruction full of violent aesthetics with scientific precision and artistic clarity, and the details are extremely realistic.", + "seed": 15307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_6670_seed_15307_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28026617302093654, + "vlm_logprob": -0.28026617302093654, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.28026617302093654, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0564175844192505 + ], + "opposite_logprob_tokens": [ + -0.4314175844192505 + ], + "logprob_reward": -1.0564175844192505, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0564175844192505, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.215239480137825, + 0.09704040735960007, + 0.06749290972948074, + 0.034853409975767136, + 0.027163559570908546, + 0.02642223797738552, + 0.015090839937329292, + 0.40028372406959534, + 0.09198448061943054, + 0.02442891150712967 + ], + "credit_map": [ + 0.03078877553343773, + 0.010328477248549461, + 0.042553890496492386, + 0.02215300127863884, + 0.031217845156788826, + 0.04242824390530586, + 0.013207572512328625, + 0.02256167121231556, + 0.011170949786901474, + 0.008432911708950996, + 0.012531490065157413, + 0.013592200353741646, + 0.014427781105041504, + 0.017395751550793648, + 0.008519851602613926, + 0.01096946932375431, + 0.01364975981414318, + 0.005006106570363045, + 0.012894811108708382, + 0.005709464196115732, + 0.0039120991714298725, + 0.012487096711993217, + 0.004304877482354641, + 0.009528696537017822, + 0.006164196878671646, + 0.0036278709303587675, + 0.005849766544997692, + 0.0061442116275429726, + 0.0037377020344138145, + 0.00367832207120955, + 0.003007191000506282, + 0.0026441491208970547, + 0.0031122704967856407, + 0.003216387238353491, + 0.0024668099358677864, + 0.0036435851361602545, + 0.0034405181650072336, + 0.004914365708827972, + 0.0026589634362608194, + 0.0037106582894921303, + 0.0030929550994187593, + 0.0025645424611866474, + 0.0021901277359575033, + 0.005315840244293213, + 0.003550372552126646, + 0.0030344172846525908, + 0.004136829636991024, + 0.0025371527299284935, + 0.0017804050585255027, + 0.0012294386979192495, + 0.0012975196586921811, + 0.001851805136539042, + 0.0014160987921059132, + 0.0021687927655875683, + 0.0019208519952371716, + 0.0034259280655533075, + 0.09009665250778198, + 0.02726784348487854, + 0.026783034205436707, + 0.02755512110888958, + 0.05704047158360481, + 0.07375361025333405, + 0.026213055476546288, + 0.07157395780086517, + 0.01428321935236454, + 0.009339539334177971, + 0.009989983402192593, + 0.015096410177648067, + 0.010071531869471073, + 0.013047939166426659, + 0.01000597421079874, + 0.010149880312383175, + 0.004157906398177147, + 0.003251610556617379, + 0.0024045344907790422, + 0.003652264131233096, + 0.0023191971704363823, + 0.0030009490437805653, + 0.0030254435259848833, + 0.0026170052587985992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -1.0564175844192505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3136765658855438 + ], + "opposite_logprob_tokens": [ + -1.3136765956878662 + ], + "logprob_reward": -0.3136765658855438, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3136765658855438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1289152354001999, + 0.11667837202548981, + 0.11478970944881439, + 0.05895882099866867, + 0.025291386991739273, + 0.05196452885866165, + 0.03384529799222946, + 0.3357737064361572, + 0.08950883150100708, + 0.044274214655160904 + ], + "credit_map": [ + 0.015651939436793327, + 0.006156768649816513, + 0.02182244136929512, + 0.019380999729037285, + 0.00984192918986082, + 0.019955815747380257, + 0.01335075218230486, + 0.02275455743074417, + 0.012604980729520321, + 0.010848353616893291, + 0.014305326156318188, + 0.020436784252524376, + 0.01795763336122036, + 0.015760796144604683, + 0.011629588901996613, + 0.013134882785379887, + 0.024601085111498833, + 0.0066932933405041695, + 0.02216164954006672, + 0.012159505859017372, + 0.006202274467796087, + 0.017446568235754967, + 0.007066881284117699, + 0.018458422273397446, + 0.008381457068026066, + 0.007096634712070227, + 0.007340273354202509, + 0.010427920147776604, + 0.008880117908120155, + 0.005319082178175449, + 0.006179099902510643, + 0.0053342231549322605, + 0.0029225877951830626, + 0.0026549939066171646, + 0.002431081607937813, + 0.0036058283876627684, + 0.003810156136751175, + 0.005016763228923082, + 0.0025832499377429485, + 0.0022667208686470985, + 0.005346386693418026, + 0.004805122967809439, + 0.004328019917011261, + 0.01124704908579588, + 0.007288813125342131, + 0.005584123078733683, + 0.008357150480151176, + 0.0050078523345291615, + 0.004659675061702728, + 0.0027685670647770166, + 0.004089692607522011, + 0.003373592160642147, + 0.0026952289044857025, + 0.004818478599190712, + 0.0032371939159929752, + 0.008202862925827503, + 0.07514145970344543, + 0.023296955972909927, + 0.02191038429737091, + 0.022552603855729103, + 0.04780643433332443, + 0.06313414126634598, + 0.02162902243435383, + 0.060302622616291046, + 0.013632137328386307, + 0.009558407589793205, + 0.010001234710216522, + 0.016235781833529472, + 0.009489314630627632, + 0.011607130989432335, + 0.009513400495052338, + 0.009471398778259754, + 0.008190909400582314, + 0.005503491964191198, + 0.00351201300509274, + 0.005848305765539408, + 0.0050196899101138115, + 0.006172680761665106, + 0.005055360961705446, + 0.004971752874553204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.3136765658855438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090303771197796 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090303771197796, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090303771197796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2691805362701416, + 0.17792829871177673, + 0.06895957142114639, + 0.05076786130666733, + 0.033500928431749344, + 0.0675097331404686, + 0.03844522684812546, + 0.1546776294708252, + 0.10045024007558823, + 0.03857996687293053 + ], + "credit_map": [ + 0.03891495242714882, + 0.011992045678198338, + 0.05486335977911949, + 0.02522820048034191, + 0.03713291510939598, + 0.052294373512268066, + 0.01725790463387966, + 0.031496766954660416, + 0.022117940708994865, + 0.01312434021383524, + 0.025294478982686996, + 0.020600629970431328, + 0.023418443277478218, + 0.04058344289660454, + 0.012893594801425934, + 0.019895438104867935, + 0.011961414478719234, + 0.0062620737589895725, + 0.010555950924754143, + 0.009474574588239193, + 0.004190544132143259, + 0.012325682677328587, + 0.004961557220667601, + 0.009227780625224113, + 0.012112983502447605, + 0.0046102507039904594, + 0.008651121519505978, + 0.005055244080722332, + 0.004300994798541069, + 0.004736416041851044, + 0.005491962190717459, + 0.0058088903315365314, + 0.0036418477538973093, + 0.0035998423118144274, + 0.003758653299883008, + 0.003793851938098669, + 0.00521383062005043, + 0.006495147477835417, + 0.0030210174154490232, + 0.003976736683398485, + 0.008701898157596588, + 0.005427850876003504, + 0.005743532907217741, + 0.013904046267271042, + 0.008974998258054256, + 0.007871395908296108, + 0.01007760688662529, + 0.006808403879404068, + 0.0048324898816645145, + 0.00362652656622231, + 0.004857588093727827, + 0.005062715616077185, + 0.0038630799390375614, + 0.0052604516968131065, + 0.004263394977897406, + 0.006678979843854904, + 0.03208646550774574, + 0.011797728948295116, + 0.008898915722966194, + 0.010359738953411579, + 0.02164580300450325, + 0.03141460195183754, + 0.01023550983518362, + 0.02823886089026928, + 0.015605813823640347, + 0.010581745766103268, + 0.011014804244041443, + 0.014398406259715557, + 0.012319258414208889, + 0.01557986717671156, + 0.009441455826163292, + 0.011508888565003872, + 0.006997525691986084, + 0.004325208719819784, + 0.0030747042037546635, + 0.0041501703672111034, + 0.005238554440438747, + 0.005867326632142067, + 0.004119035322219133, + 0.004807440564036369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011090303771197796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819971203804016 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.01819971203804016, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819971203804016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3963499069213867, + 0.313734233379364, + 0.05977894365787506, + 0.037950482219457626, + 0.027245383709669113, + 0.03234283626079559, + 0.015185041353106499, + 0.07794918119907379, + 0.01928517036139965, + 0.020178813487291336 + ], + "credit_map": [ + 0.06269429624080658, + 0.01491681206971407, + 0.06435836851596832, + 0.03821908310055733, + 0.019905662164092064, + 0.05927944555878639, + 0.045136693865060806, + 0.09183952212333679, + 0.054869599640369415, + 0.019303377717733383, + 0.04237139970064163, + 0.041560158133506775, + 0.027122875675559044, + 0.056415483355522156, + 0.022876616567373276, + 0.049214720726013184, + 0.012011862359941006, + 0.0038167801685631275, + 0.010754326358437538, + 0.010454387404024601, + 0.003610991407185793, + 0.007212156895548105, + 0.004649062640964985, + 0.007269375026226044, + 0.005768793635070324, + 0.0034674424678087234, + 0.007206196431070566, + 0.006263305898755789, + 0.0026715046260505915, + 0.0052961488254368305, + 0.003910331055521965, + 0.003366759978234768, + 0.0032389468979090452, + 0.0022720256820321083, + 0.0033472434151917696, + 0.0027952969539910555, + 0.0027329137083142996, + 0.007146422751247883, + 0.0028574324678629637, + 0.002855102065950632, + 0.003603458171710372, + 0.003503915388137102, + 0.0030110252555459738, + 0.007273344788700342, + 0.003869757056236267, + 0.0034076261799782515, + 0.004111704882234335, + 0.0035620059352368116, + 0.002685151295736432, + 0.0013239142717793584, + 0.0011172207305207849, + 0.0012270648730918765, + 0.0023652578238397837, + 0.002067156136035919, + 0.0013482444919645786, + 0.003051032079383731, + 0.01646934635937214, + 0.005326256155967712, + 0.005523523315787315, + 0.005320779979228973, + 0.010906205512583256, + 0.015230965800583363, + 0.0057020168751478195, + 0.013470089063048363, + 0.0029075678903609514, + 0.0017443860415369272, + 0.002409510314464569, + 0.0026851482689380646, + 0.0020161899738013744, + 0.003092673374339938, + 0.002051060553640127, + 0.0023786346428096294, + 0.003160631051287055, + 0.001785608590580523, + 0.0015037920093163848, + 0.002630949020385742, + 0.0026164380833506584, + 0.003624677425250411, + 0.0024412439670413733, + 0.0024154731072485447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01819971203804016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001946698990650475 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.001946698990650475, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001946698990650475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961996704339981, + 0.13074415922164917, + 0.12137110531330109, + 0.05901918560266495, + 0.04626382887363434, + 0.06393074989318848, + 0.04640206694602966, + 0.13302072882652283, + 0.13096578419208527, + 0.0720827504992485 + ], + "credit_map": [ + 0.02991276979446411, + 0.00839273165911436, + 0.036488886922597885, + 0.02226858399808407, + 0.023569174110889435, + 0.037630680948495865, + 0.014758510515093803, + 0.0231783390045166, + 0.015843363478779793, + 0.012294602580368519, + 0.01726360060274601, + 0.02200322039425373, + 0.017024165019392967, + 0.018350297585129738, + 0.012423500418663025, + 0.015541404485702515, + 0.023672381415963173, + 0.008437701500952244, + 0.022684583440423012, + 0.020996732637286186, + 0.008139513432979584, + 0.016078170388936996, + 0.009070129133760929, + 0.012291898019611835, + 0.00954893697053194, + 0.005426671821624041, + 0.007403242401778698, + 0.007332036271691322, + 0.007136763539165258, + 0.010608700104057789, + 0.0059905690141022205, + 0.005572263617068529, + 0.005590884014964104, + 0.005422264337539673, + 0.004288279451429844, + 0.005667731165885925, + 0.005933254491537809, + 0.00854573119431734, + 0.004123250488191843, + 0.00669243186712265, + 0.0072489273734390736, + 0.005255762487649918, + 0.005981144495308399, + 0.014834568835794926, + 0.007622075732797384, + 0.007326386868953705, + 0.009967673569917679, + 0.005694209132343531, + 0.007431458681821823, + 0.003372045699506998, + 0.00677641574293375, + 0.004826189950108528, + 0.004012105520814657, + 0.006439097225666046, + 0.004173590801656246, + 0.009371165186166763, + 0.028813792392611504, + 0.009447142481803894, + 0.009472656063735485, + 0.009810863994061947, + 0.0214494988322258, + 0.02480371482670307, + 0.009962956421077251, + 0.019260108470916748, + 0.020454270765185356, + 0.01402379386126995, + 0.01401737704873085, + 0.018476208671927452, + 0.015478894114494324, + 0.01968483440577984, + 0.014382553286850452, + 0.014447865076363087, + 0.012038078159093857, + 0.009897301904857159, + 0.005735591519623995, + 0.012004701420664787, + 0.007157244253903627, + 0.007718962617218494, + 0.010516040958464146, + 0.007014834322035313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.001946698990650475 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, using a bullet-time effect, captures in astonishing detail the moment an arrow completely pierces a red apple. The arrow, with a sleek black carbon fiber shaft and a sharp, gleaming steel tip, has just burst from the other side of the apple. The perfectly ripe apple, its skin a vibrant crimson, is now shattered inward by the immense force of the impact. A spectacular, three-dimensional corona of juice and pulp erupts around the arrow, with thousands of individual, crystalline droplets and tiny fibers of flesh frozen in mid-air. Some larger droplets remain perfectly spherical due to surface tension, while others are stretched into delicate threads of liquid. The entire scene is set against a pure black, non-reflective background. Sharp, bright, high-speed studio lighting illuminates every tiny detail from the side, capturing every facet of the splashing droplets, making them sparkle like countless tiny diamonds while creating crisp highlights on the smooth skin of the apple and the metal surface of the arrow. The entire picture presents an extreme surreal texture, freezing a moment of destruction full of violent aesthetics with scientific precision and artistic clarity, and the details are extremely realistic.", + "seed": 15308, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_6670_seed_15308_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10250935335643589, + "vlm_logprob": -0.10250935335643589, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.10250935335643589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2540610432624817 + ], + "opposite_logprob_tokens": [ + -1.504060983657837 + ], + "logprob_reward": -0.2540610432624817, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2540610432624817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08139891922473907, + 0.053804609924554825, + 0.22269484400749207, + 0.27237963676452637, + 0.04098387062549591, + 0.0653357282280922, + 0.06899195909500122, + 0.0644644945859909, + 0.09558797627687454, + 0.03435796499252319 + ], + "credit_map": [ + 0.01277250237762928, + 0.006147207226604223, + 0.022676371037960052, + 0.006669191177934408, + 0.008826956152915955, + 0.00552367651835084, + 0.009073475375771523, + 0.009709540754556656, + 0.00784240011125803, + 0.004724113270640373, + 0.009815914556384087, + 0.006715449504554272, + 0.007734620478004217, + 0.005109956953674555, + 0.006759383715689182, + 0.005102769937366247, + 0.03260508179664612, + 0.02073809504508972, + 0.03541496396064758, + 0.033209577202796936, + 0.014015276916325092, + 0.01804656907916069, + 0.02618555538356304, + 0.04247971624135971, + 0.01510632038116455, + 0.010594680905342102, + 0.01477957982569933, + 0.06298458576202393, + 0.011361394077539444, + 0.0155629338696599, + 0.10612744092941284, + 0.03586269170045853, + 0.00684861745685339, + 0.0020322282798588276, + 0.006207361817359924, + 0.003776739351451397, + 0.0022704971488565207, + 0.004071141593158245, + 0.008984616957604885, + 0.0067926691845059395, + 0.0077060516923666, + 0.005476924125105143, + 0.00796364713460207, + 0.010854166001081467, + 0.005613905843347311, + 0.006237259600311518, + 0.012631332501769066, + 0.008852449245750904, + 0.005673679523169994, + 0.006227143574506044, + 0.011324967257678509, + 0.010467391461133957, + 0.007138524204492569, + 0.00547958305105567, + 0.009566574357450008, + 0.013114091008901596, + 0.005179459694772959, + 0.004218770656734705, + 0.009309628047049046, + 0.007102916948497295, + 0.00463599618524313, + 0.004916648380458355, + 0.008004993200302124, + 0.021096084266901016, + 0.010124655440449715, + 0.004847863223403692, + 0.027026234194636345, + 0.01626594550907612, + 0.007990315556526184, + 0.007354167755693197, + 0.009527362883090973, + 0.01245143823325634, + 0.0029038528446108103, + 0.0021602807100862265, + 0.004358505830168724, + 0.004457227420061827, + 0.002309159841388464, + 0.0030833715572953224, + 0.004774097353219986, + 0.010311471298336983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.2540610432624817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16053208708763123 + ], + "opposite_logprob_tokens": [ + -1.9105321168899536 + ], + "logprob_reward": -0.16053208708763123, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16053208708763123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15400075912475586, + 0.11423661559820175, + 0.3377520442008972, + 0.11231067031621933, + 0.030349165201187134, + 0.04494188353419304, + 0.04716389998793602, + 0.06072019413113594, + 0.0678720474243164, + 0.03065272606909275 + ], + "credit_map": [ + 0.01919780671596527, + 0.00991730485111475, + 0.02852226048707962, + 0.014247758314013481, + 0.0122869573533535, + 0.00895130354911089, + 0.02614283747971058, + 0.03473452478647232, + 0.015606394037604332, + 0.007083934731781483, + 0.031115276739001274, + 0.02343001216650009, + 0.010723182000219822, + 0.0065267314203083515, + 0.010936612263321877, + 0.008814476430416107, + 0.04672764241695404, + 0.0172991082072258, + 0.06421008706092834, + 0.05136018618941307, + 0.01887507550418377, + 0.025219043716788292, + 0.042666833847761154, + 0.07139406353235245, + 0.007163276895880699, + 0.0055602481588721275, + 0.00789717212319374, + 0.02149137295782566, + 0.008001556620001793, + 0.008616616018116474, + 0.035865236073732376, + 0.017715195193886757, + 0.007451833225786686, + 0.0017296040896326303, + 0.0031689160969108343, + 0.0033089860808104277, + 0.0022155730985105038, + 0.0045024030841887, + 0.00474082725122571, + 0.003231021575629711, + 0.003301655873656273, + 0.003856765804812312, + 0.005368003621697426, + 0.007285750936716795, + 0.0038962382823228836, + 0.003747567068785429, + 0.010629876516759396, + 0.00685602193698287, + 0.0028422484174370766, + 0.0033588192891329527, + 0.00629477109760046, + 0.010388257913291454, + 0.004572173580527306, + 0.0033454163931310177, + 0.0070296092890203, + 0.009332604706287384, + 0.00492384796962142, + 0.0031871891114860773, + 0.011800249107182026, + 0.007341553922742605, + 0.002979637123644352, + 0.004313969518989325, + 0.008637044578790665, + 0.0175367072224617, + 0.005899244919419289, + 0.0027830873150378466, + 0.0219047199934721, + 0.013161112554371357, + 0.003519629128277302, + 0.004443018231540918, + 0.006469817832112312, + 0.009691417217254639, + 0.0027019265107810497, + 0.0018914106767624617, + 0.0040660141967237, + 0.003497819183394313, + 0.0024587027728557587, + 0.0026001071091741323, + 0.0042948671616613865, + 0.00914187915623188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.16053208708763123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370317071676254 + ], + "opposite_logprob_tokens": [ + -3.408703088760376 + ], + "logprob_reward": -0.03370317071676254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370317071676254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08617094159126282, + 0.06380246579647064, + 0.17437028884887695, + 0.15828165411949158, + 0.03577212989330292, + 0.07785779237747192, + 0.08291647583246231, + 0.09799332916736603, + 0.19895465672016144, + 0.023880261927843094 + ], + "credit_map": [ + 0.009806467220187187, + 0.007992200553417206, + 0.017819344997406006, + 0.010747174732387066, + 0.010976540856063366, + 0.006526533514261246, + 0.011772331781685352, + 0.010530347004532814, + 0.009467142634093761, + 0.004723209422081709, + 0.013117440044879913, + 0.013030746951699257, + 0.006673647556453943, + 0.004559180233627558, + 0.0071731372736394405, + 0.005057960283011198, + 0.023488318547606468, + 0.013795328326523304, + 0.023211495950818062, + 0.03125183284282684, + 0.014693054370582104, + 0.016690727323293686, + 0.02207215130329132, + 0.02916736900806427, + 0.00967088621109724, + 0.007521087769418955, + 0.011184480972588062, + 0.031144900247454643, + 0.010592962615191936, + 0.01190229132771492, + 0.05218130350112915, + 0.024083757773041725, + 0.010145594365894794, + 0.0021232531871646643, + 0.004195245448499918, + 0.0032732647377997637, + 0.002188408514484763, + 0.0037861017044633627, + 0.005322227720171213, + 0.004738034680485725, + 0.00641686050221324, + 0.007465288508683443, + 0.008699147962033749, + 0.012300961650907993, + 0.005794491618871689, + 0.007074316032230854, + 0.01610611192882061, + 0.014000620692968369, + 0.007661155425012112, + 0.005740414373576641, + 0.019467948004603386, + 0.015676449984312057, + 0.0051952386274933815, + 0.005191307049244642, + 0.010401327162981033, + 0.013582639396190643, + 0.009570463560521603, + 0.005478715058416128, + 0.012495935894548893, + 0.011434078216552734, + 0.0056621842086315155, + 0.008129038847982883, + 0.01135254930704832, + 0.03387036174535751, + 0.01773149147629738, + 0.006731478963047266, + 0.0670829489827156, + 0.040039341896772385, + 0.0067854877561330795, + 0.01269248966127634, + 0.019868919625878334, + 0.02802250348031521, + 0.0021417748648673296, + 0.0021808824967592955, + 0.0025554257445037365, + 0.0028617391362786293, + 0.002105883089825511, + 0.0027590724639594555, + 0.002799524227157235, + 0.006475958973169327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.03370317071676254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204906851053238 + ], + "opposite_logprob_tokens": [ + -2.812049150466919 + ], + "logprob_reward": -0.06204906851053238, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204906851053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742441445589066, + 0.1443178355693817, + 0.15970303118228912, + 0.07478229701519012, + 0.016651427373290062, + 0.06108797341585159, + 0.02748129703104496, + 0.1492990255355835, + 0.23453301191329956, + 0.054719675332307816 + ], + "credit_map": [ + 0.012609036639332771, + 0.0059438953176140785, + 0.009032124653458595, + 0.008114242926239967, + 0.008592311292886734, + 0.006009562406688929, + 0.010390461422502995, + 0.016732776537537575, + 0.01568257436156273, + 0.010271142236888409, + 0.028512001037597656, + 0.037756577134132385, + 0.013643418438732624, + 0.009040610864758492, + 0.019278299063444138, + 0.010133217088878155, + 0.022889502346515656, + 0.013046986423432827, + 0.02938847802579403, + 0.022078020498156548, + 0.014491351321339607, + 0.015989815816283226, + 0.01790372096002102, + 0.023915158584713936, + 0.005945986602455378, + 0.0035760647151619196, + 0.0037571373395621777, + 0.021522190421819687, + 0.003394255181774497, + 0.004456013441085815, + 0.02661689557135105, + 0.005513753741979599, + 0.001850444939918816, + 0.0012195106828585267, + 0.0017410957952961326, + 0.003704663133248687, + 0.001725906040519476, + 0.0011955712689086795, + 0.00221765274181962, + 0.00299658440053463, + 0.008827420882880688, + 0.004103281535208225, + 0.00727190962061286, + 0.014220235869288445, + 0.0044160885736346245, + 0.005419023800641298, + 0.009308026172220707, + 0.007521987892687321, + 0.0031884273048490286, + 0.002518806140869856, + 0.007188400719314814, + 0.003920843359082937, + 0.0024766905698925257, + 0.0019414493581280112, + 0.0033348172437399626, + 0.0029118608217686415, + 0.011445000767707825, + 0.010260996408760548, + 0.03598560020327568, + 0.01976243034005165, + 0.010834154672920704, + 0.011103236116468906, + 0.017986606806516647, + 0.03192099928855896, + 0.02039926126599312, + 0.009003682062029839, + 0.07377699762582779, + 0.04706590250134468, + 0.007868509739637375, + 0.012959301471710205, + 0.02175632305443287, + 0.04170301929116249, + 0.005560083780437708, + 0.00401587737724185, + 0.006285978946834803, + 0.00618195254355669, + 0.005735178478062153, + 0.005508652422577143, + 0.006157055962830782, + 0.015274890698492527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.06204906851053238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201397204771638 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201397204771638, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201397204771638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09312501549720764, + 0.07432225346565247, + 0.14978903532028198, + 0.1799106001853943, + 0.06384250521659851, + 0.0720636397600174, + 0.048583075404167175, + 0.20352044701576233, + 0.050779495388269424, + 0.06406398117542267 + ], + "credit_map": [ + 0.01315882708877325, + 0.005468881689012051, + 0.020677268505096436, + 0.008864469826221466, + 0.009305279701948166, + 0.005984498653560877, + 0.011772171594202518, + 0.017893606796860695, + 0.010204230435192585, + 0.005155154969543219, + 0.01236214954406023, + 0.01927797682583332, + 0.006184096913784742, + 0.004096379969269037, + 0.01183802355080843, + 0.0052042328752577305, + 0.02553299255669117, + 0.011705050244927406, + 0.02875697799026966, + 0.01697157323360443, + 0.008725102059543133, + 0.013227018527686596, + 0.01568048633635044, + 0.02918982319533825, + 0.010386466048657894, + 0.007103248033672571, + 0.009754149243235588, + 0.04137416183948517, + 0.007151623722165823, + 0.01029319316148758, + 0.07097673416137695, + 0.02287100814282894, + 0.016651710495352745, + 0.004569540731608868, + 0.006638620048761368, + 0.008791032247245312, + 0.0050611295737326145, + 0.005591611377894878, + 0.009209384210407734, + 0.007329469081014395, + 0.006584665738046169, + 0.006331726908683777, + 0.007569168694317341, + 0.015047085471451283, + 0.004666672553867102, + 0.006463848985731602, + 0.013709061779081821, + 0.011691398918628693, + 0.0044883121736347675, + 0.0034653698094189167, + 0.011356053873896599, + 0.008066142909228802, + 0.002916759578511119, + 0.002773452317342162, + 0.006145835388451815, + 0.009371142834424973, + 0.018115226179361343, + 0.010859920643270016, + 0.03809863328933716, + 0.021291952580213547, + 0.010428623296320438, + 0.014914200641214848, + 0.02590337209403515, + 0.06390849500894547, + 0.004516857676208019, + 0.0032210727222263813, + 0.016136478632688522, + 0.010685293935239315, + 0.0028592131566256285, + 0.0027162772603332996, + 0.004329468123614788, + 0.006314830854535103, + 0.006133485585451126, + 0.004372534342110157, + 0.009558598510921001, + 0.007082222029566765, + 0.0036627582740038633, + 0.00621418422088027, + 0.008332986384630203, + 0.01870720088481903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.002201397204771638 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, using a bullet-time effect, captures in astonishing detail the moment an arrow completely pierces a red apple. The arrow, with a sleek black carbon fiber shaft and a sharp, gleaming steel tip, has just burst from the other side of the apple. The perfectly ripe apple, its skin a vibrant crimson, is now shattered inward by the immense force of the impact. A spectacular, three-dimensional corona of juice and pulp erupts around the arrow, with thousands of individual, crystalline droplets and tiny fibers of flesh frozen in mid-air. Some larger droplets remain perfectly spherical due to surface tension, while others are stretched into delicate threads of liquid. The entire scene is set against a pure black, non-reflective background. Sharp, bright, high-speed studio lighting illuminates every tiny detail from the side, capturing every facet of the splashing droplets, making them sparkle like countless tiny diamonds while creating crisp highlights on the smooth skin of the apple and the metal surface of the arrow. The entire picture presents an extreme surreal texture, freezing a moment of destruction full of violent aesthetics with scientific precision and artistic clarity, and the details are extremely realistic.", + "seed": 15309, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_15/index_6670_seed_15309_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029118492384441197, + "vlm_logprob": -0.029118492384441197, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.029118492384441197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019388679414987564 + ], + "opposite_logprob_tokens": [ + -4.019388675689697 + ], + "logprob_reward": -0.019388679414987564, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019388679414987564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1743900030851364, + 0.15290619432926178, + 0.12730441987514496, + 0.09320314228534698, + 0.028381027281284332, + 0.08853600174188614, + 0.0480559878051281, + 0.12729501724243164, + 0.11939553916454315, + 0.0405326709151268 + ], + "credit_map": [ + 0.013984977267682552, + 0.027690257877111435, + 0.03029661998152733, + 0.03291720896959305, + 0.012939471751451492, + 0.017540596425533295, + 0.024340137839317322, + 0.01468072272837162, + 0.021351944655179977, + 0.037408169358968735, + 0.0220026858150959, + 0.020632678642868996, + 0.013131409883499146, + 0.013039864599704742, + 0.012213966809213161, + 0.013125462457537651, + 0.01654731296002865, + 0.017983147874474525, + 0.015284244902431965, + 0.013818872161209583, + 0.008794015273451805, + 0.015235412865877151, + 0.024258360266685486, + 0.015383029356598854, + 0.012549047358334064, + 0.01149789523333311, + 0.010855540633201599, + 0.014665454626083374, + 0.009813182055950165, + 0.0059843482449650764, + 0.01508354302495718, + 0.01275411807000637, + 0.004673666786402464, + 0.00359833100810647, + 0.0038516861386597157, + 0.002698575146496296, + 0.0030357169453054667, + 0.0030118434224277735, + 0.003951224498450756, + 0.0035599814727902412, + 0.01253728661686182, + 0.008624398149549961, + 0.011870724149048328, + 0.009511088952422142, + 0.014327934943139553, + 0.007487060967832804, + 0.01610514707863331, + 0.008072352036833763, + 0.007715565152466297, + 0.0049396357499063015, + 0.006632315460592508, + 0.004158204887062311, + 0.005436031613498926, + 0.011201121844351292, + 0.004897789563983679, + 0.0030753198079764843, + 0.020106563344597816, + 0.010268239304423332, + 0.031065208837389946, + 0.009133078157901764, + 0.014115916565060616, + 0.012145709246397018, + 0.019971465691924095, + 0.010488823987543583, + 0.018303168937563896, + 0.011289913207292557, + 0.010344108566641808, + 0.018108291551470757, + 0.016287732869386673, + 0.021876947954297066, + 0.008550312370061874, + 0.014635064639151096, + 0.006264562252908945, + 0.002106577390804887, + 0.004147960804402828, + 0.009135828353464603, + 0.004320137668401003, + 0.008000366389751434, + 0.004073958843946457, + 0.0024832759518176317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.019388679414987564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10048337280750275 + ], + "opposite_logprob_tokens": [ + -2.3504834175109863 + ], + "logprob_reward": -0.10048337280750275, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10048337280750275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1476438343524933, + 0.15096354484558105, + 0.07142376154661179, + 0.061891015619039536, + 0.040364623069763184, + 0.02483062632381916, + 0.03487914428114891, + 0.19680112600326538, + 0.130496546626091, + 0.14070576429367065 + ], + "credit_map": [ + 0.013839198276400566, + 0.01838207058608532, + 0.026028793305158615, + 0.02056119404733181, + 0.017448140308260918, + 0.019252166152000427, + 0.01815146394073963, + 0.01398080587387085, + 0.02163775824010372, + 0.04041149467229843, + 0.022144924849271774, + 0.020607423037290573, + 0.012620233930647373, + 0.010912766680121422, + 0.011048229411244392, + 0.01158070843666792, + 0.009903381578624249, + 0.009167398326098919, + 0.009166051633656025, + 0.007340758573263884, + 0.005416190717369318, + 0.008196519687771797, + 0.013328355737030506, + 0.00890510343015194, + 0.009807989001274109, + 0.007196365389972925, + 0.00786196906119585, + 0.008127101697027683, + 0.0056838844902813435, + 0.004498403985053301, + 0.009816345758736134, + 0.008898954838514328, + 0.00820143148303032, + 0.005380008369684219, + 0.005843169521540403, + 0.004073950927704573, + 0.003045275341719389, + 0.003953693434596062, + 0.00484056631103158, + 0.005026528611779213, + 0.004097536206245422, + 0.0035146549344062805, + 0.0032868273556232452, + 0.002257291693240404, + 0.003520228900015354, + 0.0023200286086648703, + 0.0028352763038128614, + 0.0029987813904881477, + 0.005707919131964445, + 0.003427462186664343, + 0.0038558756932616234, + 0.0027397279627621174, + 0.004190132953226566, + 0.008020970970392227, + 0.0038637954276055098, + 0.0030732599552720785, + 0.032796699553728104, + 0.013181561604142189, + 0.05359954759478569, + 0.012427003122866154, + 0.02103186771273613, + 0.017453735694289207, + 0.03294186294078827, + 0.013368853367865086, + 0.017762256786227226, + 0.014598527923226357, + 0.011777596548199654, + 0.02031242474913597, + 0.019483057782053947, + 0.023006748408079147, + 0.009579919278621674, + 0.013976014219224453, + 0.02200775034725666, + 0.008393694646656513, + 0.011797144077718258, + 0.03526295721530914, + 0.014683563262224197, + 0.024488182738423347, + 0.010657506994903088, + 0.013414964079856873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.10048337280750275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036184806376695633 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036184806376695633, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036184806376695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17364956438541412, + 0.11565545946359634, + 0.11612393707036972, + 0.11326925456523895, + 0.1120786964893341, + 0.06307798624038696, + 0.041430845856666565, + 0.1634083092212677, + 0.039516936987638474, + 0.061788998544216156 + ], + "credit_map": [ + 0.014602553099393845, + 0.028728578239679337, + 0.02821977622807026, + 0.037896644324064255, + 0.014629766345024109, + 0.01586802676320076, + 0.022838355973362923, + 0.010865867137908936, + 0.011935011483728886, + 0.020961550995707512, + 0.016560276970267296, + 0.014762384817004204, + 0.012336676008999348, + 0.012965250760316849, + 0.013395511545240879, + 0.012738797813653946, + 0.015186352655291557, + 0.015117546543478966, + 0.013691570609807968, + 0.011777433566749096, + 0.008319150656461716, + 0.014298324473202229, + 0.02356710657477379, + 0.014166444540023804, + 0.020848862826824188, + 0.013091963715851307, + 0.015396744012832642, + 0.013293649069964886, + 0.01520542986690998, + 0.01375710777938366, + 0.01094502117484808, + 0.010730473324656487, + 0.018807243555784225, + 0.013284430839121342, + 0.01445744838565588, + 0.012074780650436878, + 0.009810669347643852, + 0.014522358775138855, + 0.013155853375792503, + 0.0159659031778574, + 0.010189484804868698, + 0.005845335312187672, + 0.008172774687409401, + 0.006279612425714731, + 0.010772005654871464, + 0.005448638461530209, + 0.01092443335801363, + 0.0054457057267427444, + 0.010992412455379963, + 0.003766485955566168, + 0.010669494047760963, + 0.002926093293353915, + 0.0035359414760023355, + 0.0023592915385961533, + 0.004233712796121836, + 0.0029474112670868635, + 0.024501489475369453, + 0.015736838802695274, + 0.029917920008301735, + 0.012350012548267841, + 0.02314978651702404, + 0.02015666477382183, + 0.02103481814265251, + 0.01656079664826393, + 0.005371747072786093, + 0.0033844190184026957, + 0.004351088311523199, + 0.006801910232752562, + 0.007629844360053539, + 0.00719144381582737, + 0.002530846744775772, + 0.0022556395269930363, + 0.009268566966056824, + 0.004785175435245037, + 0.005335616879165173, + 0.007203454617410898, + 0.009671594947576523, + 0.009449268691241741, + 0.011531195603311062, + 0.004544129129499197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0036184806376695633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587211474776268 + ], + "opposite_logprob_tokens": [ + -3.895587205886841 + ], + "logprob_reward": -0.020587211474776268, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587211474776268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10553093254566193, + 0.1847926825284958, + 0.10248919576406479, + 0.07029129564762115, + 0.10604234784841537, + 0.06269268691539764, + 0.1354358196258545, + 0.027355404570698738, + 0.06519248336553574, + 0.14017713069915771 + ], + "credit_map": [ + 0.010875756852328777, + 0.015805738046765327, + 0.020035728812217712, + 0.01972491294145584, + 0.008621555753052235, + 0.008306692354381084, + 0.013691249303519726, + 0.008469299413263798, + 0.025652043521404266, + 0.046789560467004776, + 0.0286907646805048, + 0.018330032005906105, + 0.025164160877466202, + 0.01431164238601923, + 0.014229556545615196, + 0.01162493135780096, + 0.012228389270603657, + 0.015229035168886185, + 0.012282942421734333, + 0.010623408481478691, + 0.006747376639395952, + 0.011368973180651665, + 0.01857738196849823, + 0.015431690961122513, + 0.014649652875959873, + 0.009355896152555943, + 0.010005149990320206, + 0.009708338417112827, + 0.008188371546566486, + 0.006520961411297321, + 0.006211946252733469, + 0.005650978069752455, + 0.019021375104784966, + 0.016449302434921265, + 0.015288054943084717, + 0.009529679082334042, + 0.008419452235102654, + 0.012274562381207943, + 0.01375534851104021, + 0.01130457129329443, + 0.013316620141267776, + 0.007449289783835411, + 0.009825680404901505, + 0.006088755093514919, + 0.006079790182411671, + 0.007546330336481333, + 0.007090000435709953, + 0.005296227056533098, + 0.038741737604141235, + 0.009656193666160107, + 0.04053393751382828, + 0.008485108613967896, + 0.008381280116736889, + 0.01198573037981987, + 0.010324877686798573, + 0.007326943799853325, + 0.00443409476429224, + 0.002496509812772274, + 0.004538355860859156, + 0.002887753304094076, + 0.0032917058560997248, + 0.002956933341920376, + 0.004086910281330347, + 0.002663141116499901, + 0.00975351594388485, + 0.00687645748257637, + 0.00498544704169035, + 0.01031434629112482, + 0.008762113749980927, + 0.011913737282156944, + 0.005107227712869644, + 0.007479638326913118, + 0.022497709840536118, + 0.007173370104283094, + 0.007601355202496052, + 0.037240445613861084, + 0.013539300300180912, + 0.025423290207982063, + 0.015458871610462666, + 0.011242799460887909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.020587211474776268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an arrow piercing a red apple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the arrow have a black carbon fiber shaft and a sharp steel tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the apple shattered inward with a three-dimensional corona of juice and pulp erupting around the arrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting sharp, bright, and coming from the side to create sparkling highlights on the droplets and surfaces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015147175872698426 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015147175872698426, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015147175872698426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18573957681655884, + 0.2321389615535736, + 0.0792790874838829, + 0.08900176733732224, + 0.044216059148311615, + 0.04990377277135849, + 0.055953748524188995, + 0.06374573707580566, + 0.09450873732566833, + 0.10551262646913528 + ], + "credit_map": [ + 0.0206996388733387, + 0.02804679423570633, + 0.033077605068683624, + 0.027294861152768135, + 0.017990760505199432, + 0.016422292217612267, + 0.02417610026896, + 0.0180315263569355, + 0.0309447031468153, + 0.05990864709019661, + 0.036353494971990585, + 0.03330029919743538, + 0.01661854051053524, + 0.014405624009668827, + 0.02616029977798462, + 0.014447327703237534, + 0.010187280364334583, + 0.012228189036250114, + 0.00949779711663723, + 0.007694638799875975, + 0.004607790149748325, + 0.005807177629321814, + 0.014446866698563099, + 0.014809339307248592, + 0.01413156557828188, + 0.009022737853229046, + 0.010972726158797741, + 0.010554827749729156, + 0.009281085804104805, + 0.007390760816633701, + 0.01327479537576437, + 0.01437325868755579, + 0.008484655991196632, + 0.00644045090302825, + 0.005905759986490011, + 0.004747417755424976, + 0.003929320722818375, + 0.004282990470528603, + 0.006164452992379665, + 0.004261008929461241, + 0.008803021162748337, + 0.0052735167555511, + 0.0066181207075715065, + 0.004665212240070105, + 0.007277360185980797, + 0.006644511595368385, + 0.006375509314239025, + 0.004246518947184086, + 0.00910476129502058, + 0.005453485529869795, + 0.007676814217120409, + 0.004694689996540546, + 0.005725425202399492, + 0.012857158668339252, + 0.005588871892541647, + 0.0048525370657444, + 0.01097916904836893, + 0.004706122912466526, + 0.018516527488827705, + 0.006321023218333721, + 0.004694212228059769, + 0.004809415899217129, + 0.01030041929334402, + 0.0034188413992524147, + 0.016068125143647194, + 0.007610018830746412, + 0.008349113166332245, + 0.014515005052089691, + 0.013316815719008446, + 0.017294062301516533, + 0.007967803627252579, + 0.009387779049575329, + 0.017947407439351082, + 0.004882994573563337, + 0.008080964908003807, + 0.027825843542814255, + 0.009903039783239365, + 0.01847740262746811, + 0.008950624614953995, + 0.009444339200854301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0015147175872698426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_tensor([107, 107])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_tensor([107, 107])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..889a6e777bccd398e5ab9d2da6af3c1cffa2d404 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_tensor([107, 107])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle shot, captured with the exquisite quality of high-definition digital video, captures a meticulously modified Toyota Celica GT-Four race car, precariously balanced in near-stillness on a narrow, precarious bridge constructed from building blocks. The car, a bright white vehicle adorned with classic red and green racing stripes, boasts round rally lights on the front and a massive rear wing that emphasizes its performance credentials. It rests precisely in the center of the bridge, its wide rally tires barely touching the edges. The bridge, constructed entirely of a jumbled, fragile jumble of colorful plastic building blocks, buckles slightly under the car's weight, making it appear precarious. Beneath the bridge lies a clear, gurgling stream, its rounded pebbles visible at the bottom. Bathed in the warm, soft natural light of the afternoon, the scene is set against a backdrop of a typically Ghibli-esque, idyllic, and soothing countryside: rolling emerald green hills dotted with quaint farmhouses in the distance, and a vast blue sky brimming with cotton candy-like white clouds. This tense and wondrous moment seamlessly blends realistic mechanical details with a dreamlike, fairytale setting, creating a surreal atmosphere that is both suspenseful and full of whimsy. Its warm, bright tones imbue it with tranquility and a subtle tension.", + "seed": 24925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24925_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032693186029791835, + "vlm_logprob": -0.032693186029791835, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.032693186029791835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08783642202615738 + ], + "opposite_logprob_tokens": [ + -2.837836503982544 + ], + "logprob_reward": -0.08783642202615738, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08783642202615738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34777721762657166, + 0.1180926188826561, + 0.06990691274404526, + 0.04416389763355255, + 0.035334520041942596, + 0.05516037717461586, + 0.039759498089551926, + 0.08465433120727539, + 0.034778106957674026, + 0.17037253081798553 + ], + "credit_map": [ + 0.058028336614370346, + 0.08064880222082138, + 0.021452149376273155, + 0.05197134241461754, + 0.046600405126810074, + 0.03968722000718117, + 0.02959790639579296, + 0.019791046157479286, + 0.020413560792803764, + 0.02559690922498703, + 0.009509682655334473, + 0.014150294475257397, + 0.018507057800889015, + 0.011288430541753769, + 0.012309283949434757, + 0.006317399442195892, + 0.008689495734870434, + 0.013786599040031433, + 0.007406643591821194, + 0.014866050332784653, + 0.00618437584489584, + 0.008243227377533913, + 0.005842766724526882, + 0.004887755960226059, + 0.005067100282758474, + 0.009500512853264809, + 0.004602695815265179, + 0.007075076457113028, + 0.003196194302290678, + 0.005541238933801651, + 0.003693348728120327, + 0.005487728398293257, + 0.004203395452350378, + 0.004285738803446293, + 0.003950322978198528, + 0.00604827469214797, + 0.0030918039847165346, + 0.004489316139370203, + 0.0038462358061224222, + 0.005419434513896704, + 0.007072063162922859, + 0.008503252640366554, + 0.0039240457117557526, + 0.0065918187610805035, + 0.00664461450651288, + 0.006043360102921724, + 0.007705447729676962, + 0.00867577362805605, + 0.0048410529270768166, + 0.004702617879956961, + 0.005391990765929222, + 0.009024091996252537, + 0.0034964107908308506, + 0.004573290701955557, + 0.004296282306313515, + 0.0034337586257606745, + 0.01161009818315506, + 0.009426106698811054, + 0.009547023102641106, + 0.013381429016590118, + 0.012757163494825363, + 0.00994170643389225, + 0.010833012871444225, + 0.007157796528190374, + 0.005521695129573345, + 0.004727931693196297, + 0.0032067475840449333, + 0.004059081897139549, + 0.0036058921832591295, + 0.0032213982194662094, + 0.004794409964233637, + 0.005640951916575432, + 0.02427145279943943, + 0.024222737178206444, + 0.02001192606985569, + 0.03862064331769943, + 0.013058236800134182, + 0.012298800982534885, + 0.023823240771889687, + 0.014065480791032314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.08783642202615738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026374874636530876 + ], + "opposite_logprob_tokens": [ + -3.6513748168945312 + ], + "logprob_reward": -0.026374874636530876, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026374874636530876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3818281888961792, + 0.15923947095870972, + 0.06210624799132347, + 0.01854730024933815, + 0.020008450374007225, + 0.10452432930469513, + 0.022435329854488373, + 0.05965859070420265, + 0.03679375723004341, + 0.13485835492610931 + ], + "credit_map": [ + 0.06394736468791962, + 0.06495655328035355, + 0.03167445585131645, + 0.0476866252720356, + 0.0491214320063591, + 0.048544742166996, + 0.041450291872024536, + 0.03444672003388405, + 0.023705171421170235, + 0.04156063497066498, + 0.01168979424983263, + 0.020102469250559807, + 0.01895945519208908, + 0.01640315167605877, + 0.015292156487703323, + 0.011526632122695446, + 0.0074052209965884686, + 0.005561507306993008, + 0.005669951904565096, + 0.009610067121684551, + 0.007800229825079441, + 0.005839705001562834, + 0.011086381040513515, + 0.009133181534707546, + 0.002560344059020281, + 0.002276904648169875, + 0.0018541410099714994, + 0.003251573070883751, + 0.0021013368386775255, + 0.001896591274999082, + 0.0025528105907142162, + 0.0020535981748253107, + 0.002495284890756011, + 0.0022148392163217068, + 0.0019464842043817043, + 0.003335249377414584, + 0.001472165109589696, + 0.003364905249327421, + 0.0023956014774739742, + 0.0027839206159114838, + 0.01326910499483347, + 0.014659970067441463, + 0.011976271867752075, + 0.021000703796744347, + 0.011894306167960167, + 0.009434145875275135, + 0.01403496228158474, + 0.008254862390458584, + 0.0033458396792411804, + 0.002353453077375889, + 0.00267969211563468, + 0.004227599129080772, + 0.002694440307095647, + 0.0025918520987033844, + 0.0025075580924749374, + 0.0020348946563899517, + 0.007061053067445755, + 0.0071213445626199245, + 0.006998555734753609, + 0.014685210771858692, + 0.004121371544897556, + 0.0061824084259569645, + 0.009568240493535995, + 0.003920404706150293, + 0.004911864176392555, + 0.004608784336596727, + 0.003918800503015518, + 0.004633398726582527, + 0.003166500711813569, + 0.004552541300654411, + 0.004576016217470169, + 0.006425853818655014, + 0.024529391899704933, + 0.023268988355994225, + 0.013944505713880062, + 0.020933547988533974, + 0.010589255951344967, + 0.014177351258695126, + 0.01574205979704857, + 0.01167325209826231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026374874636530876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24319088459014893, + 0.08704158663749695, + 0.059542033821344376, + 0.04886637255549431, + 0.05025624483823776, + 0.07568845897912979, + 0.030184796079993248, + 0.06262240558862686, + 0.06322137266397476, + 0.2793857455253601 + ], + "credit_map": [ + 0.035405226051807404, + 0.043651338666677475, + 0.031388428062200546, + 0.02761216089129448, + 0.029484422877430916, + 0.037376586347818375, + 0.01904240809381008, + 0.019230319187045097, + 0.016075532883405685, + 0.01807825081050396, + 0.007283602841198444, + 0.011285071261227131, + 0.008726623840630054, + 0.012105722911655903, + 0.005991574376821518, + 0.0074952165596187115, + 0.008820450864732265, + 0.008640575222671032, + 0.00532294437289238, + 0.005842863116413355, + 0.0057801189832389355, + 0.010702638886868954, + 0.006100674159824848, + 0.008331771939992905, + 0.00658442685380578, + 0.005313171073794365, + 0.003907714504748583, + 0.005945263896137476, + 0.00337047828361392, + 0.006055381614714861, + 0.006119420286267996, + 0.011570518836379051, + 0.006208421196788549, + 0.0039392379112541676, + 0.005302791483700275, + 0.006601392291486263, + 0.003845830913633108, + 0.007263497915118933, + 0.006502966862171888, + 0.010592114180326462, + 0.012799118645489216, + 0.01282190065830946, + 0.0065630953758955, + 0.010386190377175808, + 0.005288340616971254, + 0.014799132011830807, + 0.005757200997322798, + 0.007273489609360695, + 0.004500979091972113, + 0.0024271784350275993, + 0.003963257186114788, + 0.004130785819143057, + 0.0023400995414704084, + 0.00308777391910553, + 0.0034710692707449198, + 0.006263657007366419, + 0.007671509403735399, + 0.0074194129556417465, + 0.006640221457928419, + 0.010261117480695248, + 0.007257216144353151, + 0.0065416921861469746, + 0.0073748924769461155, + 0.00945634488016367, + 0.009932389482855797, + 0.006552629638463259, + 0.0054570999927818775, + 0.006844716612249613, + 0.005396951921284199, + 0.007258049678057432, + 0.006759375799447298, + 0.01502016931772232, + 0.06948069483041763, + 0.05749369040131569, + 0.01767154224216938, + 0.024110935628414154, + 0.03457941114902496, + 0.027304386720061302, + 0.02551981993019581, + 0.023225294426083565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801098629832268 + ], + "opposite_logprob_tokens": [ + -3.5298011302948 + ], + "logprob_reward": -0.029801098629832268, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801098629832268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2110964059829712, + 0.061868276447057724, + 0.04411812499165535, + 0.05012980103492737, + 0.026516085490584373, + 0.18425773084163666, + 0.05300309881567955, + 0.043185118585824966, + 0.02374483272433281, + 0.3020804822444916 + ], + "credit_map": [ + 0.04340743646025658, + 0.03128920495510101, + 0.01830572634935379, + 0.013910378329455853, + 0.025341490283608437, + 0.04513820633292198, + 0.014368783682584763, + 0.019335174933075905, + 0.008496543392539024, + 0.009470587596297264, + 0.005065159406512976, + 0.009831538423895836, + 0.008441048674285412, + 0.008967312052845955, + 0.0061669363640248775, + 0.00542915053665638, + 0.00542863504961133, + 0.008591419085860252, + 0.005064842756837606, + 0.008747395128011703, + 0.0036193085834383965, + 0.004434972070157528, + 0.004307023249566555, + 0.003924529068171978, + 0.005579206161201, + 0.0066813393495976925, + 0.0052122510969638824, + 0.009877759031951427, + 0.006000032648444176, + 0.005921510048210621, + 0.006294256076216698, + 0.004563446622341871, + 0.0036246245726943016, + 0.004847974982112646, + 0.00282698473893106, + 0.0040695867501199245, + 0.0023343118373304605, + 0.003907985985279083, + 0.0025321627035737038, + 0.002372455084696412, + 0.029953278601169586, + 0.03656835854053497, + 0.01559201069176197, + 0.019894063472747803, + 0.011355533264577389, + 0.030654484406113625, + 0.01640150137245655, + 0.023838501423597336, + 0.006424637045711279, + 0.004079203121364117, + 0.007976573891937733, + 0.012677602469921112, + 0.005162813235074282, + 0.0038331958930939436, + 0.006590841803699732, + 0.006258232984691858, + 0.005605785176157951, + 0.005123490933328867, + 0.00454062782227993, + 0.006106665823608637, + 0.006686286069452763, + 0.004769281949847937, + 0.00552546139806509, + 0.004827519878745079, + 0.002743450226262212, + 0.0035710379015654325, + 0.0028819760773330927, + 0.003579719690605998, + 0.0022426333744078875, + 0.002678293501958251, + 0.0027559504378587008, + 0.0032917724456638098, + 0.045991960912942886, + 0.032849133014678955, + 0.031809721142053604, + 0.06356727331876755, + 0.02408798411488533, + 0.025093235075473785, + 0.05085166543722153, + 0.02782950922846794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.029801098629832268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31577491760253906, + 0.13611561059951782, + 0.13411951065063477, + 0.070789635181427, + 0.08288785070180893, + 0.022959353402256966, + 0.022860219702124596, + 0.028241802006959915, + 0.026838840916752815, + 0.15941229462623596 + ], + "credit_map": [ + 0.05053334683179855, + 0.0427527017891407, + 0.04761429876089096, + 0.031744662672281265, + 0.038953009992837906, + 0.05054229870438576, + 0.023228595033288002, + 0.030405987054109573, + 0.02165171690285206, + 0.03139414265751839, + 0.01122070848941803, + 0.009144305251538754, + 0.014741926454007626, + 0.027606487274169922, + 0.007419694680720568, + 0.012936633080244064, + 0.0214726310223341, + 0.028178973123431206, + 0.009018725715577602, + 0.008351576514542103, + 0.015118730254471302, + 0.030902286991477013, + 0.009084504097700119, + 0.011992083862423897, + 0.009130378253757954, + 0.0172136090695858, + 0.006120748817920685, + 0.005320683587342501, + 0.005317831877619028, + 0.015505722723901272, + 0.005728157702833414, + 0.006452508270740509, + 0.01046837866306305, + 0.010938320308923721, + 0.008994271978735924, + 0.011152905412018299, + 0.006985360290855169, + 0.016778096556663513, + 0.008781704120337963, + 0.00878881849348545, + 0.0034748914185911417, + 0.0029618723783642054, + 0.002127424580976367, + 0.002120456425473094, + 0.001910925144329667, + 0.004366893321275711, + 0.0025854201521724463, + 0.003411471378058195, + 0.0033183221239596605, + 0.003197914455085993, + 0.0026103188283741474, + 0.003722257213667035, + 0.002400708384811878, + 0.003421751782298088, + 0.002062729327008128, + 0.0021262194495648146, + 0.0035525911953300238, + 0.0038636887911707163, + 0.0034798176493495703, + 0.004591676406562328, + 0.002001448068767786, + 0.0029001610819250345, + 0.004280301742255688, + 0.003572119865566492, + 0.003927782643586397, + 0.0033555200789123774, + 0.002680064644664526, + 0.0049497331492602825, + 0.0024130470119416714, + 0.00241717416793108, + 0.0035003204829990864, + 0.003595199203118682, + 0.020027801394462585, + 0.017451655119657516, + 0.018460871651768684, + 0.04019836708903313, + 0.012861295603215694, + 0.010968099348247051, + 0.027077319100499153, + 0.012366876937448978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00525094848126173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, captured with the exquisite quality of high-definition digital video, captures a meticulously modified Toyota Celica GT-Four race car, precariously balanced in near-stillness on a narrow, precarious bridge constructed from building blocks. The car, a bright white vehicle adorned with classic red and green racing stripes, boasts round rally lights on the front and a massive rear wing that emphasizes its performance credentials. It rests precisely in the center of the bridge, its wide rally tires barely touching the edges. The bridge, constructed entirely of a jumbled, fragile jumble of colorful plastic building blocks, buckles slightly under the car's weight, making it appear precarious. Beneath the bridge lies a clear, gurgling stream, its rounded pebbles visible at the bottom. Bathed in the warm, soft natural light of the afternoon, the scene is set against a backdrop of a typically Ghibli-esque, idyllic, and soothing countryside: rolling emerald green hills dotted with quaint farmhouses in the distance, and a vast blue sky brimming with cotton candy-like white clouds. This tense and wondrous moment seamlessly blends realistic mechanical details with a dreamlike, fairytale setting, creating a surreal atmosphere that is both suspenseful and full of whimsy. Its warm, bright tones imbue it with tranquility and a subtle tension.", + "seed": 24926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24926_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02421518457122147, + "vlm_logprob": -0.02421518457122147, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02421518457122147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06888078898191452 + ], + "opposite_logprob_tokens": [ + -2.943880796432495 + ], + "logprob_reward": -0.06888078898191452, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06888078898191452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2635498046875, + 0.05537739768624306, + 0.06095774099230766, + 0.12325720489025116, + 0.09637898206710815, + 0.05308505520224571, + 0.06288984417915344, + 0.06987390667200089, + 0.1802985519170761, + 0.0343315489590168 + ], + "credit_map": [ + 0.04439922049641609, + 0.04311640188097954, + 0.02206987887620926, + 0.020203305408358574, + 0.05188795179128647, + 0.0270666666328907, + 0.029921969398856163, + 0.024884387850761414, + 0.006366182118654251, + 0.008056489750742912, + 0.004625300411134958, + 0.004926248453557491, + 0.01058176252990961, + 0.007444180082529783, + 0.008489987812936306, + 0.004887244664132595, + 0.00643271254375577, + 0.00853404588997364, + 0.006414418108761311, + 0.006403622683137655, + 0.011834791861474514, + 0.0058832187205553055, + 0.009102882817387581, + 0.006352050695568323, + 0.011394398286938667, + 0.01665930263698101, + 0.007605195045471191, + 0.01302564237266779, + 0.030313778668642044, + 0.014746378175914288, + 0.018197134137153625, + 0.011315383948385715, + 0.016261601820588112, + 0.017370885238051414, + 0.00507305609062314, + 0.0071519105695188046, + 0.020250918343663216, + 0.010703042149543762, + 0.012979608960449696, + 0.00658796401694417, + 0.005891398061066866, + 0.007615264505147934, + 0.004749702755361795, + 0.004427023231983185, + 0.012832961976528168, + 0.006164560094475746, + 0.006934529170393944, + 0.004469616338610649, + 0.005767177790403366, + 0.0076789287850260735, + 0.007146025542169809, + 0.007108398247510195, + 0.01196154486387968, + 0.00770995020866394, + 0.009035016410052776, + 0.0064828055910766125, + 0.008280365727841854, + 0.011239990592002869, + 0.003951997961848974, + 0.005553189665079117, + 0.018177861347794533, + 0.007736699655652046, + 0.00987530592828989, + 0.0050584967248141766, + 0.02474275976419449, + 0.045294493436813354, + 0.015789717435836792, + 0.01396159827709198, + 0.031498052179813385, + 0.02047019824385643, + 0.015801914036273956, + 0.012739815749228, + 0.0035143927671015263, + 0.0055213812738657, + 0.002914057346060872, + 0.002828860655426979, + 0.007783276028931141, + 0.004588854964822531, + 0.004156029783189297, + 0.0030246933456510305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.06888078898191452 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076541476882994175 + ], + "opposite_logprob_tokens": [ + -4.882654190063477 + ], + "logprob_reward": -0.0076541476882994175, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076541476882994175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.301134318113327, + 0.07918567955493927, + 0.08079828321933746, + 0.04201502352952957, + 0.06690774112939835, + 0.06608880311250687, + 0.02861660160124302, + 0.07342444360256195, + 0.22160442173480988, + 0.040224701166152954 + ], + "credit_map": [ + 0.03392393887042999, + 0.03882873058319092, + 0.05110679194331169, + 0.0284699909389019, + 0.05108904838562012, + 0.02236894704401493, + 0.0420990064740181, + 0.033247850835323334, + 0.008339392021298409, + 0.010008943267166615, + 0.0076070306822657585, + 0.008605264127254486, + 0.01285817101597786, + 0.0059504942037165165, + 0.01375304814428091, + 0.012063329108059406, + 0.0071218181401491165, + 0.009847515262663364, + 0.016959942877292633, + 0.010385787114501, + 0.008896656334400177, + 0.0065622832626104355, + 0.009653647430241108, + 0.011370630003511906, + 0.006194906309247017, + 0.005816456396132708, + 0.004234399180859327, + 0.004529272671788931, + 0.007310172077268362, + 0.003675225656479597, + 0.005424287635833025, + 0.004830304067581892, + 0.008398104459047318, + 0.012363913469016552, + 0.009176221676170826, + 0.006310645025223494, + 0.009529520757496357, + 0.0069655729457736015, + 0.007558830082416534, + 0.006604930851608515, + 0.009019502438604832, + 0.011338221840560436, + 0.012061978690326214, + 0.007144789211452007, + 0.006024855189025402, + 0.0059118447825312614, + 0.006780615542083979, + 0.007806998677551746, + 0.006140405777841806, + 0.0063437484204769135, + 0.002649616915732622, + 0.0025132864248007536, + 0.003462074790149927, + 0.0026246982160955667, + 0.0027469571214169264, + 0.002135813469067216, + 0.013006279245018959, + 0.016713930293917656, + 0.006655142176896334, + 0.005435509141534567, + 0.011260009370744228, + 0.006583963055163622, + 0.007116685155779123, + 0.0066529265604913235, + 0.027731573209166527, + 0.052469346672296524, + 0.020774925127625465, + 0.017049232497811317, + 0.040744028985500336, + 0.024233689531683922, + 0.021666269749403, + 0.016935361549258232, + 0.0066419546492397785, + 0.009395409375429153, + 0.0030772131867706776, + 0.0034558381885290146, + 0.005438066087663174, + 0.004548774566501379, + 0.003768783761188388, + 0.00389865948818624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0076541476882994175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864111352711916 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.00864111352711916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864111352711916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2763099670410156, + 0.08096639066934586, + 0.12800800800323486, + 0.0752800703048706, + 0.05140531808137894, + 0.06790149211883545, + 0.07498081773519516, + 0.050926804542541504, + 0.1612619012594223, + 0.03295915201306343 + ], + "credit_map": [ + 0.0460820272564888, + 0.048869237303733826, + 0.04138168692588806, + 0.02489706501364708, + 0.025223219767212868, + 0.0314047634601593, + 0.031433772295713425, + 0.027018195018172264, + 0.008481791242957115, + 0.009677435271441936, + 0.017171183601021767, + 0.009450444020330906, + 0.009366939775645733, + 0.006076643243432045, + 0.010702553205192089, + 0.010039391927421093, + 0.017699752002954483, + 0.02321976236999035, + 0.017562495544552803, + 0.019786590710282326, + 0.010512192733585835, + 0.009035797789692879, + 0.01390881184488535, + 0.016282614320516586, + 0.011252188123762608, + 0.017439017072319984, + 0.008371797390282154, + 0.009365986101329327, + 0.007798607926815748, + 0.007256750017404556, + 0.0055785696022212505, + 0.00821715034544468, + 0.008631465025246143, + 0.012014648877084255, + 0.0056716627441346645, + 0.005244211293756962, + 0.0058685955591499805, + 0.004768890794366598, + 0.004540790803730488, + 0.00466505391523242, + 0.01282041147351265, + 0.015027420595288277, + 0.009162220172584057, + 0.007193687837570906, + 0.005851367954164743, + 0.0062912688590586185, + 0.005445243325084448, + 0.0061098746955394745, + 0.016853881999850273, + 0.01676689274609089, + 0.007147953379899263, + 0.006751723121851683, + 0.007127139251679182, + 0.007614840753376484, + 0.005835321731865406, + 0.006883063353598118, + 0.008124913088977337, + 0.006886290851980448, + 0.005333570763468742, + 0.00755076901987195, + 0.006962544750422239, + 0.005696938373148441, + 0.005663021933287382, + 0.004708756227046251, + 0.020346371456980705, + 0.03738665208220482, + 0.0141545245423913, + 0.012843886390328407, + 0.03208158537745476, + 0.017012646421790123, + 0.015644684433937073, + 0.011791553348302841, + 0.004452244844287634, + 0.005121448542922735, + 0.0034626678097993135, + 0.003283923026174307, + 0.0050085498951375484, + 0.004854058846831322, + 0.003773486940190196, + 0.0030027690809220076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.00864111352711916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2278032749891281, + 0.07757297903299332, + 0.05291242524981499, + 0.0852724239230156, + 0.08077355474233627, + 0.06811363995075226, + 0.06538332253694534, + 0.07820164412260056, + 0.22338367998600006, + 0.040583085268735886 + ], + "credit_map": [ + 0.02858697809278965, + 0.04616450518369675, + 0.02713026851415634, + 0.016951004043221474, + 0.039100199937820435, + 0.023218411952257156, + 0.026743223890662193, + 0.019908681511878967, + 0.01494826190173626, + 0.011201035231351852, + 0.0076249693520367146, + 0.0068524861708283424, + 0.010488986037671566, + 0.007674441207200289, + 0.008521394804120064, + 0.01026140246540308, + 0.006026028655469418, + 0.006963571533560753, + 0.009724068455398083, + 0.005934765562415123, + 0.006694386247545481, + 0.004702711943536997, + 0.006500085815787315, + 0.006366811227053404, + 0.009180511347949505, + 0.015165270306169987, + 0.008364618755877018, + 0.012987208552658558, + 0.013171189464628696, + 0.008270733058452606, + 0.008491935208439827, + 0.009640957228839397, + 0.008803038857877254, + 0.012369617819786072, + 0.013554387725889683, + 0.009901197627186775, + 0.011474971659481525, + 0.00687763374298811, + 0.008820712566375732, + 0.008971993811428547, + 0.008976245298981667, + 0.012797728180885315, + 0.00922502763569355, + 0.007732875179499388, + 0.009473324753344059, + 0.006644919980317354, + 0.006735685281455517, + 0.006527839694172144, + 0.014114606194198132, + 0.013066669926047325, + 0.006792818661779165, + 0.005356954876333475, + 0.008273843675851822, + 0.006042968947440386, + 0.005282831843942404, + 0.006452621892094612, + 0.010422084480524063, + 0.015883225947618484, + 0.009370791725814342, + 0.007653593551367521, + 0.009802315384149551, + 0.009335297159850597, + 0.00896318256855011, + 0.00677115423604846, + 0.032341111451387405, + 0.05626632273197174, + 0.018032949417829514, + 0.017068099230527878, + 0.041963402181863785, + 0.023696109652519226, + 0.01983933337032795, + 0.014176342636346817, + 0.009273720905184746, + 0.008628150448203087, + 0.0029621105641126633, + 0.003730847267434001, + 0.004928667563945055, + 0.0039024772122502327, + 0.0035847348626703024, + 0.003572379471734166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.03370121121406555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198661444708705 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198661444708705, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198661444708705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27685943245887756, + 0.11795073747634888, + 0.08459881693124771, + 0.0518767312169075, + 0.022850634530186653, + 0.035299450159072876, + 0.10195646435022354, + 0.03396447002887726, + 0.1605539172887802, + 0.11408943682909012 + ], + "credit_map": [ + 0.04392686486244202, + 0.043911948800086975, + 0.04579383507370949, + 0.03382113203406334, + 0.026316946372389793, + 0.02500258758664131, + 0.023815708234906197, + 0.03427039086818695, + 0.017423585057258606, + 0.016463352367281914, + 0.01639278046786785, + 0.008667525835335255, + 0.02304457128047943, + 0.014669913798570633, + 0.009322969242930412, + 0.01196602638810873, + 0.012395933270454407, + 0.008987191133201122, + 0.017256032675504684, + 0.009308777749538422, + 0.009603500366210938, + 0.007727503776550293, + 0.009841612540185452, + 0.009478256106376648, + 0.009699199348688126, + 0.011885901913046837, + 0.005009295418858528, + 0.004298022016882896, + 0.0066181388683617115, + 0.0046190787106752396, + 0.00559038994833827, + 0.004156698938459158, + 0.002843765774741769, + 0.003654738189652562, + 0.003386309603229165, + 0.002442538971081376, + 0.0027170004323124886, + 0.002396744443103671, + 0.0030673944856971502, + 0.0023421398364007473, + 0.0036550823133438826, + 0.00498921936377883, + 0.007163283880800009, + 0.004803111311048269, + 0.003339504124596715, + 0.0031458924058824778, + 0.003597365692257881, + 0.0046059866435825825, + 0.013311335816979408, + 0.01961353048682213, + 0.011916915886104107, + 0.018167316913604736, + 0.008949617855250835, + 0.0099918432533741, + 0.007777974475175142, + 0.012227923609316349, + 0.00421524653211236, + 0.005999495275318623, + 0.003266719402745366, + 0.003802735824137926, + 0.0052932677790522575, + 0.004499966744333506, + 0.0036525072064250708, + 0.0032345294021070004, + 0.030697094276547432, + 0.02488716132938862, + 0.022336505353450775, + 0.01194907445460558, + 0.018934601917862892, + 0.019321992993354797, + 0.013093561865389347, + 0.01933389902114868, + 0.02563130110502243, + 0.02312849648296833, + 0.013159610331058502, + 0.008172763511538506, + 0.009161463938653469, + 0.014106315560638905, + 0.009465677663683891, + 0.011263793334364891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.002198661444708705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, captured with the exquisite quality of high-definition digital video, captures a meticulously modified Toyota Celica GT-Four race car, precariously balanced in near-stillness on a narrow, precarious bridge constructed from building blocks. The car, a bright white vehicle adorned with classic red and green racing stripes, boasts round rally lights on the front and a massive rear wing that emphasizes its performance credentials. It rests precisely in the center of the bridge, its wide rally tires barely touching the edges. The bridge, constructed entirely of a jumbled, fragile jumble of colorful plastic building blocks, buckles slightly under the car's weight, making it appear precarious. Beneath the bridge lies a clear, gurgling stream, its rounded pebbles visible at the bottom. Bathed in the warm, soft natural light of the afternoon, the scene is set against a backdrop of a typically Ghibli-esque, idyllic, and soothing countryside: rolling emerald green hills dotted with quaint farmhouses in the distance, and a vast blue sky brimming with cotton candy-like white clouds. This tense and wondrous moment seamlessly blends realistic mechanical details with a dreamlike, fairytale setting, creating a surreal atmosphere that is both suspenseful and full of whimsy. Its warm, bright tones imbue it with tranquility and a subtle tension.", + "seed": 24927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24927_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03318460513837636, + "vlm_logprob": -0.03318460513837636, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03318460513837636, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0759769007563591 + ], + "opposite_logprob_tokens": [ + -2.825976848602295 + ], + "logprob_reward": -0.0759769007563591, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0759769007563591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25882136821746826, + 0.0764472633600235, + 0.05275221914052963, + 0.18653886020183563, + 0.031449031084775925, + 0.08524411916732788, + 0.02902607060968876, + 0.04258716106414795, + 0.16224825382232666, + 0.07488563656806946 + ], + "credit_map": [ + 0.046249572187662125, + 0.040909089148044586, + 0.019049841910600662, + 0.023893967270851135, + 0.0398055799305439, + 0.028839746490120888, + 0.030225258320569992, + 0.029848305508494377, + 0.013866973109543324, + 0.012090989388525486, + 0.007663670927286148, + 0.0069574592635035515, + 0.01070089265704155, + 0.00863210391253233, + 0.007938907481729984, + 0.00859626941382885, + 0.011841094121336937, + 0.008039013482630253, + 0.004576980136334896, + 0.003993198275566101, + 0.007073068991303444, + 0.008722458966076374, + 0.0036025680601596832, + 0.004903838038444519, + 0.060663852840662, + 0.06239062547683716, + 0.01081040408462286, + 0.00908569898456335, + 0.011553730815649033, + 0.010889208875596523, + 0.008863385766744614, + 0.012281956151127815, + 0.003951873630285263, + 0.003838171483948827, + 0.003868753556162119, + 0.0032472372986376286, + 0.004934971686452627, + 0.004667815752327442, + 0.00326355523429811, + 0.0036766543053090572, + 0.012734008021652699, + 0.009319856762886047, + 0.0061400835402309895, + 0.010756809264421463, + 0.013668853789567947, + 0.009653032757341862, + 0.014980870299041271, + 0.00799060333520174, + 0.004077342338860035, + 0.004491383675485849, + 0.002890986390411854, + 0.0035576033405959606, + 0.0045166159979999065, + 0.0033677173778414726, + 0.003183552995324135, + 0.0029408675618469715, + 0.005920249968767166, + 0.005070087034255266, + 0.0047300285659730434, + 0.004387280438095331, + 0.007169128395617008, + 0.004690306726843119, + 0.005484074354171753, + 0.005136002786457539, + 0.03071296028792858, + 0.027745911851525307, + 0.01378930825740099, + 0.012880044989287853, + 0.02153805084526539, + 0.02509300597012043, + 0.01322785671800375, + 0.017261117696762085, + 0.014903304167091846, + 0.010879607871174812, + 0.005000013392418623, + 0.006381883751600981, + 0.007442682050168514, + 0.014222003519535065, + 0.006225922144949436, + 0.009830217808485031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0759769007563591 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029832571744918823 + ], + "opposite_logprob_tokens": [ + -3.529832601547241 + ], + "logprob_reward": -0.029832571744918823, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029832571744918823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14481456577777863, + 0.1001032143831253, + 0.050514571368694305, + 0.12901178002357483, + 0.021794799715280533, + 0.07572940737009048, + 0.028970377519726753, + 0.04775344952940941, + 0.2596298158168793, + 0.14167805016040802 + ], + "credit_map": [ + 0.01396435871720314, + 0.02006179466843605, + 0.01166823972016573, + 0.016104301437735558, + 0.03266819566488266, + 0.015498840250074863, + 0.016530942171812057, + 0.018317893147468567, + 0.011555776931345463, + 0.012331950478255749, + 0.006480607204139233, + 0.0087388064712286, + 0.024767089635133743, + 0.010752404108643532, + 0.014857384376227856, + 0.010619187727570534, + 0.0068317013792693615, + 0.008084634318947792, + 0.0034461223985999823, + 0.0059944214299321175, + 0.009837917983531952, + 0.005873135756701231, + 0.0056442818604409695, + 0.004802352283149958, + 0.04090576246380806, + 0.03847292438149452, + 0.011843515560030937, + 0.007776204962283373, + 0.008880511857569218, + 0.006602290086448193, + 0.00549682043492794, + 0.009033759124577045, + 0.002846683608368039, + 0.0026191554497927427, + 0.0037246719002723694, + 0.0027142567560076714, + 0.0018001296557486057, + 0.0030040356796234846, + 0.002769452752545476, + 0.002316413912922144, + 0.011788883246481419, + 0.009086362086236477, + 0.005121410358697176, + 0.007547734770923853, + 0.013960362412035465, + 0.009297775104641914, + 0.01160705927759409, + 0.0073198238387703896, + 0.004906462039798498, + 0.003781292587518692, + 0.0033282290678471327, + 0.0022226180881261826, + 0.004599167965352535, + 0.004573708865791559, + 0.0028049126267433167, + 0.0027539855800569057, + 0.005591980181634426, + 0.00625389302149415, + 0.004098755307495594, + 0.005578480660915375, + 0.006686226464807987, + 0.008216573856770992, + 0.005853030830621719, + 0.005474506877362728, + 0.06672430038452148, + 0.0664624497294426, + 0.018286440521478653, + 0.012817350216209888, + 0.025625990703701973, + 0.03513336926698685, + 0.016506783664226532, + 0.018073121085762978, + 0.04306231066584587, + 0.029370257630944252, + 0.00787819642573595, + 0.008531705476343632, + 0.006051552016288042, + 0.013955710455775261, + 0.009212174452841282, + 0.023616144433617592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.029832571744918823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644422516226768 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644422516226768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644422516226768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0903685986995697, + 0.027291100472211838, + 0.024402664974331856, + 0.08954653888940811, + 0.02035292610526085, + 0.040577277541160583, + 0.020252639427781105, + 0.06055399775505066, + 0.2532171607017517, + 0.37343713641166687 + ], + "credit_map": [ + 0.012193541042506695, + 0.010374273173511028, + 0.010994145646691322, + 0.014028895646333694, + 0.01227023359388113, + 0.009420203045010567, + 0.01074195932596922, + 0.01034533977508545, + 0.003062129719182849, + 0.0037713993806391954, + 0.002589206676930189, + 0.002674306510016322, + 0.004406624473631382, + 0.00318907224573195, + 0.0035952290054410696, + 0.0040031312964856625, + 0.003430749289691448, + 0.002846328541636467, + 0.0019463010830804706, + 0.002617362653836608, + 0.004143919795751572, + 0.00293352990411222, + 0.0037548597902059555, + 0.0027296124026179314, + 0.030643532052636147, + 0.028115496039390564, + 0.004745753947645426, + 0.003794345771893859, + 0.004287297371774912, + 0.006254867184907198, + 0.00409807451069355, + 0.007607162464410067, + 0.0030830770265311003, + 0.0026897594798356295, + 0.0022680447436869144, + 0.0020126989111304283, + 0.002945767715573311, + 0.0030233520083129406, + 0.002347624395042658, + 0.001982599264010787, + 0.006556531880050898, + 0.004261433146893978, + 0.002919320948421955, + 0.004595170263200998, + 0.006538506597280502, + 0.004999121185392141, + 0.006837015971541405, + 0.003870174288749695, + 0.0030159386806190014, + 0.003132238984107971, + 0.0015892037190496922, + 0.0018684023525565863, + 0.0028957128524780273, + 0.0029557684902101755, + 0.002589813433587551, + 0.0022055613808333874, + 0.01154018472880125, + 0.008963203057646751, + 0.003930781036615372, + 0.0039014958310872316, + 0.006774096749722958, + 0.010507135652005672, + 0.005927898455411196, + 0.00900919921696186, + 0.06929494440555573, + 0.072181835770607, + 0.014775266870856285, + 0.01271389052271843, + 0.020090369507670403, + 0.0299383457750082, + 0.0186005849391222, + 0.015621906146407127, + 0.11002667248249054, + 0.07139404118061066, + 0.016229495406150818, + 0.018840443342924118, + 0.019691547378897667, + 0.0409783273935318, + 0.025342388078570366, + 0.07093419134616852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008644422516226768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486503504216671 + ], + "opposite_logprob_tokens": [ + -3.0486502647399902 + ], + "logprob_reward": -0.0486503504216671, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486503504216671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03134816884994507, + 0.011883258819580078, + 0.020457452163100243, + 0.07260473072528839, + 0.010650732554495335, + 0.015196746215224266, + 0.019990138709545135, + 0.04010231792926788, + 0.3163851201534271, + 0.46138128638267517 + ], + "credit_map": [ + 0.006117598619312048, + 0.0032004916574805975, + 0.0031998681370168924, + 0.004301863256841898, + 0.0037318277172744274, + 0.0035090853925794363, + 0.0036490552593022585, + 0.0036383820697665215, + 0.0013877969468012452, + 0.0015196254244074225, + 0.0012538264272734523, + 0.0012378513347357512, + 0.002439983421936631, + 0.001127709518186748, + 0.0015215729363262653, + 0.0013948949053883553, + 0.0035722581669688225, + 0.0026571417693048716, + 0.0015125507488846779, + 0.0020556969102472067, + 0.00293217902071774, + 0.0032575028017163277, + 0.0019875052385032177, + 0.002482619369402528, + 0.02516823820769787, + 0.0214579738676548, + 0.005495599005371332, + 0.0034413891844451427, + 0.0034807007759809494, + 0.00489142956212163, + 0.0031175483018159866, + 0.005551854148507118, + 0.0018243738450109959, + 0.0014500730903819203, + 0.0008518249960616231, + 0.0010697779944166541, + 0.001966260140761733, + 0.0013501893263310194, + 0.0010985244298353791, + 0.001039709895849228, + 0.0023515138309448957, + 0.0015426158206537366, + 0.0016848193481564522, + 0.0018179250182583928, + 0.002843680325895548, + 0.0014093851204961538, + 0.0023408089764416218, + 0.001205997890792787, + 0.003373043844476342, + 0.0026554944925010204, + 0.0011853825999423862, + 0.0014818513300269842, + 0.003222570288926363, + 0.003874054178595543, + 0.0018558955052867532, + 0.0023418490309268236, + 0.007534387521445751, + 0.00632299343124032, + 0.002419087802991271, + 0.0024002797435969114, + 0.004647955764085054, + 0.00772955222055316, + 0.003556623123586178, + 0.005491440650075674, + 0.08823592215776443, + 0.09331066906452179, + 0.01762736216187477, + 0.015783658251166344, + 0.02409527637064457, + 0.03535112738609314, + 0.023455288261175156, + 0.018525823950767517, + 0.13585089147090912, + 0.08796525001525879, + 0.020116116851568222, + 0.02372721955180168, + 0.023230526596307755, + 0.05068790540099144, + 0.03139523044228554, + 0.08840817213058472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0486503504216671 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028187802527099848 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.0028187802527099848, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028187802527099848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2554231882095337, + 0.06802812218666077, + 0.13531026244163513, + 0.11165745556354523, + 0.0632958933711052, + 0.05357378348708153, + 0.0376679003238678, + 0.04810607433319092, + 0.17675314843654633, + 0.05018414556980133 + ], + "credit_map": [ + 0.06392055004835129, + 0.044557441025972366, + 0.0176238976418972, + 0.027558131143450737, + 0.02163553237915039, + 0.035445742309093475, + 0.01904565840959549, + 0.025636248290538788, + 0.016542762517929077, + 0.009714363142848015, + 0.006526866462081671, + 0.0060723270289599895, + 0.006097417324781418, + 0.00965824257582426, + 0.006131509318947792, + 0.007284641731530428, + 0.031202349811792374, + 0.0342855341732502, + 0.007141951005905867, + 0.010760845616459846, + 0.014961051754653454, + 0.018605049699544907, + 0.00877405982464552, + 0.009579428471624851, + 0.039429061114788055, + 0.03647637739777565, + 0.004814928397536278, + 0.004997662268579006, + 0.0064733996987342834, + 0.0072617302648723125, + 0.00489764055237174, + 0.007306666579097509, + 0.010752687230706215, + 0.008735411800444126, + 0.008581265807151794, + 0.0071636768989264965, + 0.006345728877931833, + 0.007090547122061253, + 0.00799319427460432, + 0.006633386947214603, + 0.0078013245947659016, + 0.007066305726766586, + 0.008124168030917645, + 0.006730644032359123, + 0.008257860317826271, + 0.006335864309221506, + 0.005529982969164848, + 0.003727639326825738, + 0.006075110752135515, + 0.0046514710411429405, + 0.004504477139562368, + 0.004731432069092989, + 0.005706001538783312, + 0.005035800859332085, + 0.0035409382544457912, + 0.003422672627493739, + 0.005947100464254618, + 0.005306958686560392, + 0.008644905872642994, + 0.005142380483448505, + 0.005647895857691765, + 0.006323092617094517, + 0.005356746260076761, + 0.005736998748034239, + 0.050815973430871964, + 0.055414699018001556, + 0.009767393581569195, + 0.008229516446590424, + 0.014621078968048096, + 0.01774120144546032, + 0.010842179879546165, + 0.009321118704974651, + 0.00866725854575634, + 0.006447043269872665, + 0.005469499155879021, + 0.004020312335342169, + 0.004845133051276207, + 0.01277897134423256, + 0.004236183594912291, + 0.0037197466008365154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0028187802527099848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot, captured with the exquisite quality of high-definition digital video, captures a meticulously modified Toyota Celica GT-Four race car, precariously balanced in near-stillness on a narrow, precarious bridge constructed from building blocks. The car, a bright white vehicle adorned with classic red and green racing stripes, boasts round rally lights on the front and a massive rear wing that emphasizes its performance credentials. It rests precisely in the center of the bridge, its wide rally tires barely touching the edges. The bridge, constructed entirely of a jumbled, fragile jumble of colorful plastic building blocks, buckles slightly under the car's weight, making it appear precarious. Beneath the bridge lies a clear, gurgling stream, its rounded pebbles visible at the bottom. Bathed in the warm, soft natural light of the afternoon, the scene is set against a backdrop of a typically Ghibli-esque, idyllic, and soothing countryside: rolling emerald green hills dotted with quaint farmhouses in the distance, and a vast blue sky brimming with cotton candy-like white clouds. This tense and wondrous moment seamlessly blends realistic mechanical details with a dreamlike, fairytale setting, creating a surreal atmosphere that is both suspenseful and full of whimsy. Its warm, bright tones imbue it with tranquility and a subtle tension.", + "seed": 24928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_2/index_107_seed_24928_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030464502424001692, + "vlm_logprob": -0.030464502424001692, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.030464502424001692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09106787294149399 + ], + "opposite_logprob_tokens": [ + -2.8410677909851074 + ], + "logprob_reward": -0.09106787294149399, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09106787294149399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17345057427883148, + 0.05223352462053299, + 0.04909713938832283, + 0.057599734514951706, + 0.12826696038246155, + 0.19337783753871918, + 0.09795179963111877, + 0.08989778906106949, + 0.12205582857131958, + 0.036068812012672424 + ], + "credit_map": [ + 0.03952941298484802, + 0.028320766985416412, + 0.016534237191081047, + 0.014076488092541695, + 0.015958717092871666, + 0.026072952896356583, + 0.013523436151444912, + 0.019434550777077675, + 0.00997535977512598, + 0.008475159294903278, + 0.004518528468906879, + 0.003804293228313327, + 0.00863014068454504, + 0.004261122550815344, + 0.005737920291721821, + 0.006831000559031963, + 0.007957551628351212, + 0.010250302962958813, + 0.004376993048936129, + 0.00470286700874567, + 0.008181276731193066, + 0.0057820784859359264, + 0.003809690708294511, + 0.004036379512399435, + 0.009003582410514355, + 0.00883767381310463, + 0.00575362890958786, + 0.00465562054887414, + 0.011862854473292828, + 0.005767541937530041, + 0.005592653527855873, + 0.006126172374933958, + 0.019205229356884956, + 0.014795202761888504, + 0.009788685478270054, + 0.011392060667276382, + 0.02956877090036869, + 0.013066086918115616, + 0.017380570992827415, + 0.013070356100797653, + 0.025919727981090546, + 0.02419789880514145, + 0.018231825903058052, + 0.0161046851426363, + 0.0366685725748539, + 0.025753671303391457, + 0.018208108842372894, + 0.02829335257411003, + 0.010690470226109028, + 0.009068944491446018, + 0.013166660442948341, + 0.010019822046160698, + 0.01335781067609787, + 0.012779124081134796, + 0.011499535292387009, + 0.017369434237480164, + 0.010722089558839798, + 0.012457337230443954, + 0.011351305060088634, + 0.009385628625750542, + 0.012282856740057468, + 0.010325617156922817, + 0.009953908622264862, + 0.01341904979199171, + 0.013478044420480728, + 0.009669316001236439, + 0.01820601522922516, + 0.020227180793881416, + 0.013167878612875938, + 0.013481151312589645, + 0.012191122397780418, + 0.02163511887192726, + 0.005118006374686956, + 0.003999305889010429, + 0.004539349116384983, + 0.003880639560520649, + 0.004975234158337116, + 0.0046956720761954784, + 0.003835572861135006, + 0.005025032442063093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.09106787294149399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060624770820141 + ], + "opposite_logprob_tokens": [ + -3.895606279373169 + ], + "logprob_reward": -0.02060624770820141, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060624770820141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20807987451553345, + 0.07314010709524155, + 0.034439876675605774, + 0.05633917450904846, + 0.10108315199613571, + 0.16122058033943176, + 0.11621227115392685, + 0.16346187889575958, + 0.041930846869945526, + 0.04409234598278999 + ], + "credit_map": [ + 0.043082691729068756, + 0.02870880626142025, + 0.024470757693052292, + 0.024536896497011185, + 0.013566006906330585, + 0.021797509863972664, + 0.020604079589247704, + 0.03131309896707535, + 0.016426758840680122, + 0.015424647368490696, + 0.006157787516713142, + 0.008180910721421242, + 0.008016116917133331, + 0.006226670928299427, + 0.005998734850436449, + 0.006708478555083275, + 0.006888741627335548, + 0.005839606747031212, + 0.0029677546117454767, + 0.0037368598859757185, + 0.0046176728792488575, + 0.003007215913385153, + 0.0033828490413725376, + 0.003999172709882259, + 0.008726791478693485, + 0.009059478528797626, + 0.005063078831881285, + 0.006430838722735643, + 0.00953727588057518, + 0.0055290088057518005, + 0.004684325773268938, + 0.00730836670845747, + 0.012872726656496525, + 0.013192645274102688, + 0.00796471070498228, + 0.011138997972011566, + 0.01960010826587677, + 0.00972300861030817, + 0.01733582839369774, + 0.009255116805434227, + 0.021306509152054787, + 0.02113703452050686, + 0.013930197805166245, + 0.013488449156284332, + 0.029310857877135277, + 0.021767830476164818, + 0.019318150356411934, + 0.020961526781320572, + 0.015215439721941948, + 0.019201941788196564, + 0.009311927482485771, + 0.010492317378520966, + 0.022800136357545853, + 0.010892879217863083, + 0.0173264779150486, + 0.01097114197909832, + 0.01951395906507969, + 0.01966976560652256, + 0.013561959378421307, + 0.01547536812722683, + 0.03272133693099022, + 0.01691998355090618, + 0.03019747883081436, + 0.015402011573314667, + 0.005238592624664307, + 0.005918101407587528, + 0.004579093307256699, + 0.004050024319440126, + 0.007228431757539511, + 0.004746980033814907, + 0.0055442191660404205, + 0.00462539866566658, + 0.0064291078597307205, + 0.007169308140873909, + 0.003312767716124654, + 0.004408028908073902, + 0.0067461165599524975, + 0.004682399332523346, + 0.006522162351757288, + 0.0048224483616650105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.02060624770820141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24733984470367432, + 0.14157430827617645, + 0.10676612704992294, + 0.0665232315659523, + 0.07033894956111908, + 0.08489245176315308, + 0.05420009791851044, + 0.08062031865119934, + 0.08644131571054459, + 0.06130335107445717 + ], + "credit_map": [ + 0.0426047183573246, + 0.032298922538757324, + 0.028227217495441437, + 0.02611619047820568, + 0.02115984447300434, + 0.026887938380241394, + 0.0359659269452095, + 0.034079089760780334, + 0.029116835445165634, + 0.024617284536361694, + 0.012262177653610706, + 0.014032640494406223, + 0.016668232157826424, + 0.01005147397518158, + 0.019398096948862076, + 0.01542755775153637, + 0.01879684440791607, + 0.020983228459954262, + 0.0122979199513793, + 0.009995492175221443, + 0.006982123013585806, + 0.01219742838293314, + 0.009025133214890957, + 0.016487959772348404, + 0.008493008092045784, + 0.0072272419929504395, + 0.006842137314379215, + 0.00761650875210762, + 0.012591682374477386, + 0.006368330679833889, + 0.007748828269541264, + 0.009635495021939278, + 0.007631373591721058, + 0.008242773823440075, + 0.0067468006163835526, + 0.007460360415279865, + 0.010521117597818375, + 0.009798044338822365, + 0.01090917643159628, + 0.009029301814734936, + 0.010608899407088757, + 0.013393676839768887, + 0.0071086096577346325, + 0.007069216575473547, + 0.01740495301783085, + 0.00837134663015604, + 0.011253147386014462, + 0.00968259759247303, + 0.005853754933923483, + 0.006178047973662615, + 0.00580202229321003, + 0.007069334387779236, + 0.00981099996715784, + 0.00574150774627924, + 0.008126962929964066, + 0.005617463495582342, + 0.010291761718690395, + 0.012567637488245964, + 0.00739179877564311, + 0.007835364900529385, + 0.014716029167175293, + 0.007952787913382053, + 0.012863182462751865, + 0.007001754827797413, + 0.010254655964672565, + 0.009264902211725712, + 0.010438056662678719, + 0.008124158717691898, + 0.01402298640459776, + 0.007906006649136543, + 0.01056645717471838, + 0.015864089131355286, + 0.00766865536570549, + 0.008380197919905186, + 0.006320744287222624, + 0.006803764961659908, + 0.011125342920422554, + 0.005284339189529419, + 0.00948001816868782, + 0.00624028779566288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008645604364573956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23131029307842255, + 0.09375914931297302, + 0.08817021548748016, + 0.06858308613300323, + 0.0858110710978508, + 0.09884406626224518, + 0.05203142389655113, + 0.0653662160038948, + 0.17861203849315643, + 0.03751241788268089 + ], + "credit_map": [ + 0.02505735494196415, + 0.024769717827439308, + 0.030912643298506737, + 0.0343940407037735, + 0.026470733806490898, + 0.021785089746117592, + 0.03069627657532692, + 0.03722447156906128, + 0.014676202088594437, + 0.00971931591629982, + 0.011384782381355762, + 0.013329754583537579, + 0.012419614940881729, + 0.007060551084578037, + 0.011346825398504734, + 0.013822113163769245, + 0.01526598073542118, + 0.017422780394554138, + 0.010274483822286129, + 0.007708818186074495, + 0.006835730280727148, + 0.009085210040211678, + 0.008930453099310398, + 0.012646765448153019, + 0.01107692625373602, + 0.01382496114820242, + 0.006478443741798401, + 0.00934415590018034, + 0.012396045960485935, + 0.006238357163965702, + 0.004272442776709795, + 0.0049517592415213585, + 0.010383454151451588, + 0.010615072213113308, + 0.009594923816621304, + 0.008201627060770988, + 0.016779690980911255, + 0.009131496772170067, + 0.008867228403687477, + 0.01223758328706026, + 0.012353899888694286, + 0.009426181204617023, + 0.017511440441012383, + 0.01573391817510128, + 0.016056863591074944, + 0.009304603561758995, + 0.00841071642935276, + 0.010046451352536678, + 0.006288295146077871, + 0.004693337716162205, + 0.009594002738595009, + 0.00844240840524435, + 0.005527092143893242, + 0.006327923387289047, + 0.00547926127910614, + 0.005679110065102577, + 0.007364862598478794, + 0.006742632482200861, + 0.010042035020887852, + 0.007225430104881525, + 0.009463897906243801, + 0.008660529740154743, + 0.0061004068702459335, + 0.009766428731381893, + 0.017638912424445152, + 0.014103714376688004, + 0.02875283733010292, + 0.031966645270586014, + 0.0159545186907053, + 0.017303068190813065, + 0.01667609252035618, + 0.036216262727975845, + 0.00473122950643301, + 0.0042800260707736015, + 0.0038729356601834297, + 0.003891215194016695, + 0.007848096080124378, + 0.004509121645241976, + 0.004821694456040859, + 0.003558106953278184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modified Toyota Celica GT-Four race car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car bright white with red and green racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car balanced on a narrow bridge made of colorful plastic building blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from a static low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a countryside with rolling green hills, farmhouses, and a blue sky with white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199375070631504 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.002199375070631504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199375070631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32926145195961, + 0.1175609827041626, + 0.08338253200054169, + 0.08406046032905579, + 0.08105722814798355, + 0.06437014043331146, + 0.06959334015846252, + 0.07278938591480255, + 0.06250584870576859, + 0.03541864827275276 + ], + "credit_map": [ + 0.07629276067018509, + 0.05178523436188698, + 0.02841152437031269, + 0.02574186958372593, + 0.040377676486968994, + 0.041176047176122665, + 0.030614608898758888, + 0.03486174717545509, + 0.015974948182702065, + 0.015215346589684486, + 0.010916872881352901, + 0.014706104062497616, + 0.016517359763383865, + 0.012173596769571304, + 0.017818668857216835, + 0.014238095842301846, + 0.015109175816178322, + 0.0159266646951437, + 0.008014658465981483, + 0.008470647037029266, + 0.009234710596501827, + 0.009048360399901867, + 0.007915209047496319, + 0.009663102217018604, + 0.011429264210164547, + 0.010931328870356083, + 0.011334759183228016, + 0.009750626981258392, + 0.009631182998418808, + 0.009800413623452187, + 0.006917438004165888, + 0.014265436679124832, + 0.009148876182734966, + 0.006964533589780331, + 0.011705338023602962, + 0.008858248591423035, + 0.010598286986351013, + 0.00933642778545618, + 0.011435998603701591, + 0.013009520247578621, + 0.006942671723663807, + 0.006350290030241013, + 0.008155117742717266, + 0.00850409921258688, + 0.010819260962307453, + 0.006204420235008001, + 0.008074196986854076, + 0.009320083074271679, + 0.008393123745918274, + 0.005969907622784376, + 0.009297681041061878, + 0.00807791855186224, + 0.011177463456988335, + 0.008726252242922783, + 0.006332147400826216, + 0.0116188433021307, + 0.008751343004405499, + 0.007422745227813721, + 0.008999517187476158, + 0.010642977431416512, + 0.007953685708343983, + 0.009001095779240131, + 0.0062238494865596294, + 0.01379417721182108, + 0.006645195186138153, + 0.007185321766883135, + 0.008855422027409077, + 0.010930314660072327, + 0.009673651307821274, + 0.005809026770293713, + 0.007900266908109188, + 0.005506649613380432, + 0.004207173362374306, + 0.0034035230055451393, + 0.004034474957734346, + 0.004815863911062479, + 0.004370477516204119, + 0.004220243543386459, + 0.0043389881029725075, + 0.006027904339134693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.002199375070631504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_tensor([1254, 1254])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_tensor([1254, 1254])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3222bd4d13f8ed750723fef61e5932a04c3ae730 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_tensor([1254, 1254])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic aerial drone shot slowly and steadily follows an eagle as it glides leisurely through the air. The eagle's massive wings are fully extended, each dark brown feather gleaming metallically in the golden morning sunlight, its edges sharply defined by the light. Its piercing yellow eyes gaze intently at the landscape below, its curved beak a look of resolve and power. Beneath it, a vast expanse of mountains stretches, its ridgeline undulating like the back of a giant dragon. Deep green coniferous forests cover most of the slopes, punctuated by exposed gray rock cliffs. A light morning mist lingers in the valley, adding a touch of mystery to the magnificent landscape. The eagle barely flaps its wings, skillfully riding the rising thermals in the valley, adjusting its flight with only subtle tilts of its body and the gentle waving of its tail feathers. Its flight is smooth and effortless. The sunlight shines down at a low angle, not only outlining a circle of bright outline light on the eagle's back, but also casting long, soft blue shadows on the back of the mountains. The whole picture presents an atmosphere of grandeur, tranquility and primitive vitality, full of awe of the power and freedom of nature.", + "seed": 32193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32193_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09341984745115042, + "vlm_logprob": -0.09341984745115042, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.09341984745115042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2263832688331604 + ], + "opposite_logprob_tokens": [ + -1.6013832092285156 + ], + "logprob_reward": -0.2263832688331604, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2263832688331604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924989610910416, + 0.04481035843491554, + 0.130874902009964, + 0.13716210424900055, + 0.06276316940784454, + 0.0768374353647232, + 0.0580824613571167, + 0.026746347546577454, + 0.037683505564928055, + 0.3457897901535034 + ], + "credit_map": [ + 0.0054690102115273476, + 0.005133541766554117, + 0.014857270754873753, + 0.01431762333959341, + 0.006377402693033218, + 0.006508177611976862, + 0.01106509193778038, + 0.015521779656410217, + 0.0030540383886545897, + 0.0035264224279671907, + 0.006695123855024576, + 0.010007601231336594, + 0.00400129659101367, + 0.0064371260814368725, + 0.005069319158792496, + 0.006019430235028267, + 0.009899641387164593, + 0.0158467385917902, + 0.020660536363720894, + 0.02670660801231861, + 0.009541409090161324, + 0.012738962657749653, + 0.016955895349383354, + 0.018525104969739914, + 0.010204797610640526, + 0.029296398162841797, + 0.03932154178619385, + 0.018955405801534653, + 0.00780772790312767, + 0.007552643772214651, + 0.013427052646875381, + 0.01059652864933014, + 0.007138376589864492, + 0.011363006196916103, + 0.009289560839533806, + 0.008982738479971886, + 0.005336297210305929, + 0.0062746708281338215, + 0.0057778614573180676, + 0.008600657805800438, + 0.005645389202982187, + 0.006236594170331955, + 0.011752614751458168, + 0.014351514168083668, + 0.005653166677802801, + 0.011160753667354584, + 0.008646606467664242, + 0.013390802778303623, + 0.004843757022172213, + 0.0056789047084748745, + 0.008836111053824425, + 0.01591610535979271, + 0.003462156979367137, + 0.00403178296983242, + 0.009876876138150692, + 0.005436768755316734, + 0.003260186640545726, + 0.002450635889545083, + 0.004051359836012125, + 0.005407084710896015, + 0.002457468770444393, + 0.0027070629876106977, + 0.0025396186392754316, + 0.003872930072247982, + 0.003545203246176243, + 0.0027085368055850267, + 0.006174452602863312, + 0.005529713351279497, + 0.005461756139993668, + 0.0059359874576330185, + 0.0032926180865615606, + 0.005035236477851868, + 0.02851789817214012, + 0.029194623231887817, + 0.05528196692466736, + 0.030041025951504707, + 0.04239238053560257, + 0.07194146513938904, + 0.04309501871466637, + 0.04532541334629059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.2263832688331604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995338141918182 + ], + "opposite_logprob_tokens": [ + -1.8049533367156982 + ], + "logprob_reward": -0.17995338141918182, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995338141918182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06427576392889023, + 0.09601140767335892, + 0.098661869764328, + 0.14631985127925873, + 0.11314286291599274, + 0.02993653528392315, + 0.0769335925579071, + 0.10903917998075485, + 0.10892947763204575, + 0.15674945712089539 + ], + "credit_map": [ + 0.007163413800299168, + 0.006144448183476925, + 0.01085498183965683, + 0.011313443072140217, + 0.005711182951927185, + 0.00534116942435503, + 0.011092249304056168, + 0.00665487814694643, + 0.0054505132138729095, + 0.008875438012182713, + 0.015717200934886932, + 0.029553314670920372, + 0.0062282513827085495, + 0.011982354335486889, + 0.009503847919404507, + 0.008700490929186344, + 0.006215664092451334, + 0.01216590590775013, + 0.015284954570233822, + 0.017635749652981758, + 0.009202488698065281, + 0.011997457593679428, + 0.01326459925621748, + 0.012895050458610058, + 0.01487752329558134, + 0.025062335655093193, + 0.022390617057681084, + 0.02145121805369854, + 0.01319827139377594, + 0.01580202393233776, + 0.013481620699167252, + 0.020056242123246193, + 0.010842881165444851, + 0.027708325535058975, + 0.03755553066730499, + 0.01525061298161745, + 0.005503952503204346, + 0.0043549383990466595, + 0.005003149621188641, + 0.006923476234078407, + 0.003063865238800645, + 0.002629338065162301, + 0.004934067837893963, + 0.005731640849262476, + 0.002680029021576047, + 0.0036066609900444746, + 0.003421724308282137, + 0.0038692071102559566, + 0.006640082225203514, + 0.009555968455970287, + 0.010435005649924278, + 0.01396805141121149, + 0.006997330579906702, + 0.007939349859952927, + 0.010929673910140991, + 0.010468128137290478, + 0.015574264340102673, + 0.01327007170766592, + 0.01605120860040188, + 0.02480495721101761, + 0.006287476047873497, + 0.009902161546051502, + 0.008015298284590244, + 0.015133745968341827, + 0.013399681076407433, + 0.010256909765303135, + 0.018086660653352737, + 0.01682993955910206, + 0.010133788920938969, + 0.014651073142886162, + 0.008966842666268349, + 0.016604572534561157, + 0.009145581163465977, + 0.010657343082129955, + 0.03230442851781845, + 0.015184355899691582, + 0.013803156092762947, + 0.02423059195280075, + 0.016679512336850166, + 0.034744493663311005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.17995338141918182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.399791955947876, + 0.09546399861574173, + 0.05611063167452812, + 0.06651441752910614, + 0.03254276514053345, + 0.052957627922296524, + 0.05874073505401611, + 0.021939214318990707, + 0.026602888479828835, + 0.18933571875095367 + ], + "credit_map": [ + 0.04112117737531662, + 0.028874574229121208, + 0.05949879065155983, + 0.09529732167720795, + 0.0279826819896698, + 0.03329441696405411, + 0.05254335328936577, + 0.061179615557193756, + 0.01034246664494276, + 0.007776298094540834, + 0.016669631004333496, + 0.023030655458569527, + 0.005069751758128405, + 0.01033022440969944, + 0.010075679048895836, + 0.01216929592192173, + 0.0044579957611858845, + 0.009789776988327503, + 0.008273890241980553, + 0.008257630281150341, + 0.005619237665086985, + 0.006335271056741476, + 0.005695371888577938, + 0.0076814559288322926, + 0.009895209223031998, + 0.013884959742426872, + 0.012453479692339897, + 0.00898499321192503, + 0.003981399349868298, + 0.0038922580424696207, + 0.00582078518345952, + 0.00760133471339941, + 0.0028870822861790657, + 0.004502363037317991, + 0.009582883678376675, + 0.007083785254508257, + 0.0017535981023684144, + 0.0017778028268367052, + 0.002128747757524252, + 0.0028265027794986963, + 0.004426578991115093, + 0.005240670405328274, + 0.008103565312922001, + 0.00834216270595789, + 0.005353290122002363, + 0.0066321673803031445, + 0.007551341783255339, + 0.007307854946702719, + 0.014954679645597935, + 0.004945713095366955, + 0.008418690413236618, + 0.007020869292318821, + 0.002790753496810794, + 0.004197600297629833, + 0.003530803369358182, + 0.012881623581051826, + 0.0034781871363520622, + 0.003056265413761139, + 0.002827753545716405, + 0.0037910465616732836, + 0.0013068397529423237, + 0.0026046272832900286, + 0.0017086690058931708, + 0.0031658257357776165, + 0.003532234113663435, + 0.002231365069746971, + 0.0043926239013671875, + 0.004560071509331465, + 0.0029436007607728243, + 0.0031841453164815903, + 0.0022136869374662638, + 0.0035451631993055344, + 0.012589155696332455, + 0.01610354334115982, + 0.02597852610051632, + 0.0178078506141901, + 0.02289297990500927, + 0.04198271408677101, + 0.02341245301067829, + 0.02856849692761898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01419412437826395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21315933763980865, + 0.07664719969034195, + 0.04319169744849205, + 0.13355915248394012, + 0.0971451997756958, + 0.09489021450281143, + 0.08097825199365616, + 0.038095444440841675, + 0.07461544126272202, + 0.14771811664104462 + ], + "credit_map": [ + 0.01416346151381731, + 0.02412467449903488, + 0.034324802458286285, + 0.048121146857738495, + 0.021453896537423134, + 0.015589830465614796, + 0.030491892248392105, + 0.024889620020985603, + 0.005125871393829584, + 0.0070711104199290276, + 0.011796073988080025, + 0.016839513555169106, + 0.005487952847033739, + 0.012148660607635975, + 0.00980608444660902, + 0.008371925912797451, + 0.003415606217458844, + 0.004466941114515066, + 0.006079479120671749, + 0.007395009510219097, + 0.004365865606814623, + 0.0061978609301149845, + 0.0048525636084377766, + 0.006418371573090553, + 0.010923403315246105, + 0.026783844456076622, + 0.04442209005355835, + 0.018365321680903435, + 0.00613203551620245, + 0.006613141857087612, + 0.009320561774075031, + 0.010998747311532497, + 0.008045696653425694, + 0.02632502093911171, + 0.03333088755607605, + 0.013409611769020557, + 0.004554773215204477, + 0.004036043304949999, + 0.0032501271925866604, + 0.004193032160401344, + 0.007369992323219776, + 0.008060772903263569, + 0.015798548236489296, + 0.015178033150732517, + 0.009755169041454792, + 0.012593543156981468, + 0.014017678797245026, + 0.01211647316813469, + 0.009470626711845398, + 0.011337882839143276, + 0.011624286882579327, + 0.018114931881427765, + 0.005026121158152819, + 0.008722599595785141, + 0.007598364260047674, + 0.00908343493938446, + 0.005181917455047369, + 0.0044182403944432735, + 0.005617398303002119, + 0.008784511126577854, + 0.0024956446141004562, + 0.004189458675682545, + 0.0030794143676757812, + 0.004328856244683266, + 0.007828975096344948, + 0.007021798286587, + 0.01092622708529234, + 0.010873128660023212, + 0.007738798391073942, + 0.011263101361691952, + 0.007344426587224007, + 0.011618982069194317, + 0.00958322361111641, + 0.011567322537302971, + 0.023322466760873795, + 0.012405158020555973, + 0.017228424549102783, + 0.030199702829122543, + 0.01828232780098915, + 0.02512948401272297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.023288307711482048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280154913663864 + ], + "opposite_logprob_tokens": [ + -3.773280143737793 + ], + "logprob_reward": -0.023280154913663864, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280154913663864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17937514185905457, + 0.09441809356212616, + 0.07760683447122574, + 0.10914591699838638, + 0.0881895199418068, + 0.04760275408625603, + 0.08841816335916519, + 0.07657311856746674, + 0.07437966018915176, + 0.16429080069065094 + ], + "credit_map": [ + 0.014617962762713432, + 0.016178937628865242, + 0.028040815144777298, + 0.036843132227659225, + 0.022000547498464584, + 0.015469895675778389, + 0.02618321217596531, + 0.020040633156895638, + 0.00915836263448, + 0.0104950787499547, + 0.014744952321052551, + 0.017966417595744133, + 0.005483971908688545, + 0.01055535115301609, + 0.01071851048618555, + 0.015295451506972313, + 0.005368042271584272, + 0.010858163237571716, + 0.011094201356172562, + 0.015716442838311195, + 0.006773259490728378, + 0.00921392161399126, + 0.009352035820484161, + 0.009230763651430607, + 0.00878765620291233, + 0.025811022147536278, + 0.03652510046958923, + 0.01442134939134121, + 0.004639558028429747, + 0.005985611584037542, + 0.006953824777156115, + 0.00602179579436779, + 0.008515208028256893, + 0.012525436468422413, + 0.026316657662391663, + 0.016524987295269966, + 0.006420765537768602, + 0.005571381654590368, + 0.005409957375377417, + 0.00690512266010046, + 0.005896968767046928, + 0.004858195316046476, + 0.006245913449674845, + 0.007925947196781635, + 0.005094439256936312, + 0.004669261164963245, + 0.006024071015417576, + 0.006887960713356733, + 0.009674697183072567, + 0.00969729945063591, + 0.01201611664146185, + 0.02191099151968956, + 0.006162990350276232, + 0.009806070476770401, + 0.010746179148554802, + 0.008403819054365158, + 0.013133988715708256, + 0.009778257459402084, + 0.01096218079328537, + 0.01349987555295229, + 0.004749993793666363, + 0.007339022122323513, + 0.006917026825249195, + 0.010192778892815113, + 0.008173094131052494, + 0.005589223466813564, + 0.010529196821153164, + 0.013850299641489983, + 0.008541553281247616, + 0.008878998458385468, + 0.007275816984474659, + 0.011541486717760563, + 0.01116116438060999, + 0.014435804449021816, + 0.022370675578713417, + 0.016184573993086815, + 0.019520875066518784, + 0.03748982399702072, + 0.01980634033679962, + 0.023321542888879776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.023280154913663864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic aerial drone shot slowly and steadily follows an eagle as it glides leisurely through the air. The eagle's massive wings are fully extended, each dark brown feather gleaming metallically in the golden morning sunlight, its edges sharply defined by the light. Its piercing yellow eyes gaze intently at the landscape below, its curved beak a look of resolve and power. Beneath it, a vast expanse of mountains stretches, its ridgeline undulating like the back of a giant dragon. Deep green coniferous forests cover most of the slopes, punctuated by exposed gray rock cliffs. A light morning mist lingers in the valley, adding a touch of mystery to the magnificent landscape. The eagle barely flaps its wings, skillfully riding the rising thermals in the valley, adjusting its flight with only subtle tilts of its body and the gentle waving of its tail feathers. Its flight is smooth and effortless. The sunlight shines down at a low angle, not only outlining a circle of bright outline light on the eagle's back, but also casting long, soft blue shadows on the back of the mountains. The whole picture presents an atmosphere of grandeur, tranquility and primitive vitality, full of awe of the power and freedom of nature.", + "seed": 32194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32194_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06583896856755019, + "vlm_logprob": -0.06583896856755019, + "vlm_yes_no_margin": 2.949999976158142, + "vlm_token_credit": -0.06583896856755019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08044232428073883 + ], + "opposite_logprob_tokens": [ + -2.580442428588867 + ], + "logprob_reward": -0.08044232428073883, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08044232428073883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18663491308689117, + 0.06988310813903809, + 0.11225268244743347, + 0.06674323230981827, + 0.3008545935153961, + 0.028599323704838753, + 0.058237336575984955, + 0.08093536645174026, + 0.0378701277077198, + 0.05798930674791336 + ], + "credit_map": [ + 0.021527478471398354, + 0.014592777006328106, + 0.027476372197270393, + 0.016944969072937965, + 0.03262101858854294, + 0.018341170623898506, + 0.032069697976112366, + 0.02306145615875721, + 0.008951476775109768, + 0.0076452456414699554, + 0.008416245691478252, + 0.0053160483948886395, + 0.011808431707322598, + 0.008659100160002708, + 0.008561964146792889, + 0.010524602606892586, + 0.01561980415135622, + 0.012944486923515797, + 0.020049232989549637, + 0.008295470848679543, + 0.00938525889068842, + 0.014482078142464161, + 0.013539622537791729, + 0.017936747521162033, + 0.007789868861436844, + 0.01003564614802599, + 0.0094551220536232, + 0.007928811945021152, + 0.009934595786035061, + 0.0084229139611125, + 0.007677151821553707, + 0.00549912778660655, + 0.03662528470158577, + 0.051334574818611145, + 0.03923730552196503, + 0.04321140795946121, + 0.022327227517962456, + 0.025655463337898254, + 0.042750369757413864, + 0.03971298038959503, + 0.0036848345771431923, + 0.002242026384919882, + 0.0038178248796612024, + 0.0028829884249716997, + 0.0027806449215859175, + 0.0046936823055148125, + 0.0035968422889709473, + 0.00490048062056303, + 0.00629869569092989, + 0.00498130451887846, + 0.007747507654130459, + 0.007721917703747749, + 0.004806687124073505, + 0.005131847690790892, + 0.014426658861339092, + 0.007122717797756195, + 0.010484321042895317, + 0.009746288880705833, + 0.01122083980590105, + 0.00872090831398964, + 0.016519954428076744, + 0.007337469607591629, + 0.009609741158783436, + 0.007295849733054638, + 0.005467545241117477, + 0.003343332326039672, + 0.007345969323068857, + 0.003475232981145382, + 0.002953031798824668, + 0.008228926919400692, + 0.003128656418994069, + 0.003927433397620916, + 0.00876531284302473, + 0.007563362829387188, + 0.006187891587615013, + 0.004134975839406252, + 0.006092551164329052, + 0.007121931295841932, + 0.007969152182340622, + 0.010154134593904018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.08044232428073883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428595632314682 + ], + "opposite_logprob_tokens": [ + -2.01785945892334 + ], + "logprob_reward": -0.1428595632314682, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428595632314682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28383609652519226, + 0.07716508209705353, + 0.10068706423044205, + 0.07099507004022598, + 0.12456320226192474, + 0.035732269287109375, + 0.061769645661115646, + 0.09064367413520813, + 0.06653366982936859, + 0.08807422965765 + ], + "credit_map": [ + 0.04213520511984825, + 0.021205971017479897, + 0.038966357707977295, + 0.024296622723340988, + 0.058059949427843094, + 0.03317758068442345, + 0.03219753876328468, + 0.03379686549305916, + 0.009166169911623001, + 0.009980718605220318, + 0.010789699852466583, + 0.007333934307098389, + 0.010726667009294033, + 0.009564097039401531, + 0.009685547091066837, + 0.00991824734956026, + 0.014274761080741882, + 0.012456727214157581, + 0.015783734619617462, + 0.008202753029763699, + 0.009685784578323364, + 0.011445662938058376, + 0.014315865002572536, + 0.014521765522658825, + 0.009944867342710495, + 0.011403689160943031, + 0.010335502214729786, + 0.009436448104679585, + 0.00871263351291418, + 0.0071073779836297035, + 0.007060650736093521, + 0.006993895396590233, + 0.01625332236289978, + 0.021947525441646576, + 0.015393964014947414, + 0.01619541458785534, + 0.014506914652884007, + 0.01199132576584816, + 0.014306927099823952, + 0.01396779716014862, + 0.004347949754446745, + 0.006219619885087013, + 0.005126472096890211, + 0.0035339435562491417, + 0.003959104418754578, + 0.003447454422712326, + 0.004049175418913364, + 0.005048547405749559, + 0.005708850920200348, + 0.00558308744803071, + 0.008747064508497715, + 0.011763937771320343, + 0.004953582771122456, + 0.005602259188890457, + 0.01152106188237667, + 0.007889802567660809, + 0.007199916057288647, + 0.012826875783503056, + 0.012960425578057766, + 0.014517934061586857, + 0.014065766707062721, + 0.008770032785832882, + 0.011696357280015945, + 0.008606360293924809, + 0.006113455165177584, + 0.003620048053562641, + 0.019463298842310905, + 0.005016252864152193, + 0.004224858246743679, + 0.017200728878378868, + 0.005914678331464529, + 0.004980340600013733, + 0.01420286763459444, + 0.011458300054073334, + 0.010583304800093174, + 0.006794314365833998, + 0.008914312347769737, + 0.012137927114963531, + 0.01053442433476448, + 0.013448772951960564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.1428595632314682 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864376410841942 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864376410841942, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864376410841942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1898527890443802, + 0.034064289182424545, + 0.2638823091983795, + 0.04854431375861168, + 0.22991958260536194, + 0.026985719799995422, + 0.027780119329690933, + 0.03942447900772095, + 0.054596997797489166, + 0.08494941145181656 + ], + "credit_map": [ + 0.024473872035741806, + 0.013866536319255829, + 0.0303910244256258, + 0.017514703795313835, + 0.03348338603973389, + 0.018154101446270943, + 0.02895425073802471, + 0.023014914244413376, + 0.004053490702062845, + 0.0036914709489792585, + 0.005021106451749802, + 0.00365880667231977, + 0.005559105426073074, + 0.003132416633889079, + 0.005191116593778133, + 0.0037567750550806522, + 0.0326983742415905, + 0.03331774100661278, + 0.04091676324605942, + 0.0215462613850832, + 0.02269754745066166, + 0.0346025712788105, + 0.037033356726169586, + 0.04106970876455307, + 0.006747044622898102, + 0.006784758996218443, + 0.006219273898750544, + 0.003979858942329884, + 0.008968989364802837, + 0.006859054788947105, + 0.004676339216530323, + 0.004308993928134441, + 0.04730236530303955, + 0.023891158401966095, + 0.030540931969881058, + 0.010435190051794052, + 0.025173425674438477, + 0.042064063251018524, + 0.015769222751259804, + 0.03474322706460953, + 0.00386061891913414, + 0.002656917553395033, + 0.004101615399122238, + 0.0020893823821097612, + 0.0025905005168169737, + 0.0029782771598547697, + 0.0034803098533302546, + 0.005228097550570965, + 0.003598499810323119, + 0.003100058762356639, + 0.0034796481486409903, + 0.002822937909513712, + 0.002592995995655656, + 0.003490127855911851, + 0.00454963743686676, + 0.004146212711930275, + 0.0069451951421797276, + 0.003959969151765108, + 0.007023119367659092, + 0.003351416438817978, + 0.004860630724579096, + 0.0045907460153102875, + 0.004746499937027693, + 0.003946901299059391, + 0.009091859683394432, + 0.005854046903550625, + 0.006470305845141411, + 0.0076746148988604546, + 0.0056990692391991615, + 0.007318240590393543, + 0.005867542698979378, + 0.006621317472308874, + 0.010756684467196465, + 0.009238217957317829, + 0.009739686734974384, + 0.006236211396753788, + 0.007982979528605938, + 0.010509329847991467, + 0.013747199438512325, + 0.016739102080464363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.04864376410841942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305906221270561 + ], + "opposite_logprob_tokens": [ + -3.1680591106414795 + ], + "logprob_reward": -0.04305906221270561, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305906221270561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10136830806732178, + 0.05089890584349632, + 0.22476954758167267, + 0.0964607447385788, + 0.20354370772838593, + 0.0704217329621315, + 0.052576757967472076, + 0.09417130053043365, + 0.04505773261189461, + 0.06073125824332237 + ], + "credit_map": [ + 0.01063755713403225, + 0.008335706777870655, + 0.017149237915873528, + 0.010540958493947983, + 0.01498816255480051, + 0.009941876865923405, + 0.017135754227638245, + 0.012639052234590054, + 0.005628873594105244, + 0.006939716637134552, + 0.006622414570301771, + 0.005059271585196257, + 0.006602282635867596, + 0.005207937676459551, + 0.006918438710272312, + 0.007919970899820328, + 0.02711609937250614, + 0.023458868265151978, + 0.041922811418771744, + 0.016260024160146713, + 0.019580606371164322, + 0.030873889103531837, + 0.030420614406466484, + 0.035136640071868896, + 0.011816172860562801, + 0.012302896939218044, + 0.012929698452353477, + 0.007332721725106239, + 0.016022417694330215, + 0.01485525444149971, + 0.011133921332657337, + 0.010067658498883247, + 0.033287063241004944, + 0.019704826176166534, + 0.0300728902220726, + 0.014862305484712124, + 0.02377181313931942, + 0.039933737367391586, + 0.015408025123178959, + 0.026503048837184906, + 0.010832779109477997, + 0.00567706348374486, + 0.01088343933224678, + 0.00571388378739357, + 0.005929135717451572, + 0.010450346395373344, + 0.010119087062776089, + 0.01081599947065115, + 0.006932235788553953, + 0.005565247032791376, + 0.007152882870286703, + 0.004038648214191198, + 0.005834224168211222, + 0.00849324930459261, + 0.006662575528025627, + 0.007897695526480675, + 0.015496782027184963, + 0.010459449142217636, + 0.017086097970604897, + 0.010165700688958168, + 0.009241999126970768, + 0.010823528282344341, + 0.011225530877709389, + 0.009672218933701515, + 0.007361035328358412, + 0.00409215223044157, + 0.006678150966763496, + 0.004439704120159149, + 0.00534761231392622, + 0.008536865934729576, + 0.003497711382806301, + 0.005104499403387308, + 0.007483791559934616, + 0.0065821330063045025, + 0.008797001093626022, + 0.005608898121863604, + 0.00599697744473815, + 0.006920842919498682, + 0.008916603401303291, + 0.010425010696053505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.04305906221270561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419012900441885 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.01419012900441885, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419012900441885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2019449919462204, + 0.04484128579497337, + 0.20129196345806122, + 0.06435324996709824, + 0.23765437304973602, + 0.04649785906076431, + 0.02615262381732464, + 0.052489813417196274, + 0.056828197091817856, + 0.06794556230306625 + ], + "credit_map": [ + 0.023152096197009087, + 0.01389129925519228, + 0.030513832345604897, + 0.01608489453792572, + 0.039222825318574905, + 0.02169281616806984, + 0.029628876596689224, + 0.02775835245847702, + 0.005813332740217447, + 0.00438151927664876, + 0.008278892375528812, + 0.002797541208565235, + 0.005204301327466965, + 0.00433448376134038, + 0.00723237544298172, + 0.00679883873090148, + 0.025347895920276642, + 0.02475137449800968, + 0.028017953038215637, + 0.017177904024720192, + 0.017640482634305954, + 0.02759833075106144, + 0.03059082105755806, + 0.030167192220687866, + 0.00601186603307724, + 0.00926310196518898, + 0.007531342562288046, + 0.006478242576122284, + 0.01055834162980318, + 0.00823887623846531, + 0.007626373786479235, + 0.008645111694931984, + 0.029219668358564377, + 0.03001352585852146, + 0.029687905684113503, + 0.018223384395241737, + 0.031826965510845184, + 0.0525892898440361, + 0.020203806459903717, + 0.025889841839671135, + 0.007337442133575678, + 0.0040168799459934235, + 0.006589694879949093, + 0.0038156670052558184, + 0.0046857320703566074, + 0.008193672634661198, + 0.005355944391340017, + 0.006502826698124409, + 0.003512617666274309, + 0.002728468971326947, + 0.0032134477514773607, + 0.0025645531713962555, + 0.0024353715125471354, + 0.0031278096139431, + 0.004028316121548414, + 0.004542037844657898, + 0.005631482228636742, + 0.008135010488331318, + 0.005450021475553513, + 0.005192279350012541, + 0.008097488433122635, + 0.006372326985001564, + 0.0074675497598946095, + 0.006143651902675629, + 0.0037451633252203465, + 0.0025703618302941322, + 0.019107013940811157, + 0.0033299948554486036, + 0.0028386313933879137, + 0.018973052501678467, + 0.0031858531292527914, + 0.003078127047047019, + 0.007220457773655653, + 0.0069852909073233604, + 0.008170941844582558, + 0.004937885794788599, + 0.00621639471501112, + 0.011047573760151863, + 0.010191396810114384, + 0.013175619766116142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.01419012900441885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic aerial drone shot slowly and steadily follows an eagle as it glides leisurely through the air. The eagle's massive wings are fully extended, each dark brown feather gleaming metallically in the golden morning sunlight, its edges sharply defined by the light. Its piercing yellow eyes gaze intently at the landscape below, its curved beak a look of resolve and power. Beneath it, a vast expanse of mountains stretches, its ridgeline undulating like the back of a giant dragon. Deep green coniferous forests cover most of the slopes, punctuated by exposed gray rock cliffs. A light morning mist lingers in the valley, adding a touch of mystery to the magnificent landscape. The eagle barely flaps its wings, skillfully riding the rising thermals in the valley, adjusting its flight with only subtle tilts of its body and the gentle waving of its tail feathers. Its flight is smooth and effortless. The sunlight shines down at a low angle, not only outlining a circle of bright outline light on the eagle's back, but also casting long, soft blue shadows on the back of the mountains. The whole picture presents an atmosphere of grandeur, tranquility and primitive vitality, full of awe of the power and freedom of nature.", + "seed": 32195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32195_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03220192310400307, + "vlm_logprob": -0.03220192310400307, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03220192310400307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009980772621929646 + ], + "opposite_logprob_tokens": [ + -4.75998067855835 + ], + "logprob_reward": -0.009980772621929646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009980772621929646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05350382253527641, + 0.17971095442771912, + 0.18056543171405792, + 0.06828202307224274, + 0.19021673500537872, + 0.05099724978208542, + 0.10836415737867355, + 0.02480083890259266, + 0.08738293498754501, + 0.0561758428812027 + ], + "credit_map": [ + 0.0035873122978955507, + 0.005312815774232149, + 0.006342570763081312, + 0.010881739668548107, + 0.005565354600548744, + 0.007464799098670483, + 0.006786057259887457, + 0.00756317051127553, + 0.014960934408009052, + 0.015728725120425224, + 0.02357339672744274, + 0.03752754628658295, + 0.024689745157957077, + 0.018369048833847046, + 0.022808877751231194, + 0.022052675485610962, + 0.015196925960481167, + 0.01726028323173523, + 0.025572018697857857, + 0.02810586430132389, + 0.019287124276161194, + 0.027843840420246124, + 0.026035666465759277, + 0.02126369997859001, + 0.005625277291983366, + 0.009780188091099262, + 0.010605870746076107, + 0.008275989443063736, + 0.007884114980697632, + 0.01340717263519764, + 0.006615334656089544, + 0.006088074296712875, + 0.008271541446447372, + 0.008007872849702835, + 0.05711645260453224, + 0.009118863381445408, + 0.014747432433068752, + 0.06084350124001503, + 0.02127937413752079, + 0.010831690393388271, + 0.0051983315497636795, + 0.004439204931259155, + 0.007456034421920776, + 0.006762971635907888, + 0.007561941165477037, + 0.006708154920488596, + 0.008900289423763752, + 0.003970325458794832, + 0.010377634316682816, + 0.01172376424074173, + 0.01700928807258606, + 0.013671892695128918, + 0.01688806526362896, + 0.013357107527554035, + 0.008954128250479698, + 0.01638227514922619, + 0.002496616682037711, + 0.003433284116908908, + 0.003148674964904785, + 0.0017531538615003228, + 0.00397251034155488, + 0.003931833431124687, + 0.0029791942797601223, + 0.0030855725053697824, + 0.015312070958316326, + 0.02977132610976696, + 0.005117096472531557, + 0.00332849589176476, + 0.004025010857731104, + 0.0046838815324008465, + 0.006061641499400139, + 0.019083412364125252, + 0.004194919019937515, + 0.00771133229136467, + 0.0104698920622468, + 0.005998199339956045, + 0.0068803210742771626, + 0.006768942344933748, + 0.005796477198600769, + 0.008355759084224701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.009980772621929646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906727284193039 + ], + "opposite_logprob_tokens": [ + -2.464067220687866 + ], + "logprob_reward": -0.08906727284193039, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906727284193039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06154278665781021, + 0.23191934823989868, + 0.07613956183195114, + 0.1731610745191574, + 0.27154287695884705, + 0.01976458728313446, + 0.057928990572690964, + 0.014918603003025055, + 0.029130300506949425, + 0.06395193934440613 + ], + "credit_map": [ + 0.005611744709312916, + 0.006093642208725214, + 0.008043996058404446, + 0.009379773400723934, + 0.005146229173988104, + 0.006481367629021406, + 0.009998605586588383, + 0.010787432081997395, + 0.0214383602142334, + 0.023098312318325043, + 0.03009776584804058, + 0.04547635093331337, + 0.030462313443422318, + 0.02263263240456581, + 0.028312671929597855, + 0.030400944873690605, + 0.006351142656058073, + 0.005429502576589584, + 0.010225022211670876, + 0.01279646996408701, + 0.007655169349163771, + 0.011441381648182869, + 0.011941291391849518, + 0.010299580171704292, + 0.010112948715686798, + 0.015385034494102001, + 0.02527291141450405, + 0.03378673270344734, + 0.025751197710633278, + 0.024085644632577896, + 0.02537965029478073, + 0.013386942446231842, + 0.011772855184972286, + 0.009375070221722126, + 0.08326172828674316, + 0.011209472082555294, + 0.021373173221945763, + 0.08640323579311371, + 0.03281744197010994, + 0.01532988715916872, + 0.0015575458528473973, + 0.001526492997072637, + 0.002970839384943247, + 0.003723608097061515, + 0.0021574213169515133, + 0.002801398979499936, + 0.002585941692814231, + 0.002441338263452053, + 0.00567115331068635, + 0.006731318775564432, + 0.00871457252651453, + 0.007125569507479668, + 0.009865029715001583, + 0.006711027584969997, + 0.00539073022082448, + 0.007719585672020912, + 0.0013554003089666367, + 0.001531721674837172, + 0.0023182877339422703, + 0.0015171620761975646, + 0.0020623463205993176, + 0.003474090015515685, + 0.001351896207779646, + 0.001307699247263372, + 0.004497735295444727, + 0.00843697041273117, + 0.00259293126873672, + 0.0018889571074396372, + 0.0019709430634975433, + 0.00214762962423265, + 0.0021982593461871147, + 0.005396872293204069, + 0.004996975883841515, + 0.014333209954202175, + 0.007542370352894068, + 0.005377165507525206, + 0.005799446254968643, + 0.005519604776054621, + 0.00497503811493516, + 0.015408127568662167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.08906727284193039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019375376868993044 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019375376868993044, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019375376868993044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11395741254091263, + 0.06889435648918152, + 0.10167046636343002, + 0.10940784215927124, + 0.393603652715683, + 0.029399394989013672, + 0.033957432955503464, + 0.060117632150650024, + 0.01971745304763317, + 0.06927435845136642 + ], + "credit_map": [ + 0.010362837463617325, + 0.009920504875481129, + 0.014704212546348572, + 0.009577222168445587, + 0.010909494012594223, + 0.02363951876759529, + 0.016329927369952202, + 0.018513696268200874, + 0.007898759096860886, + 0.008892298676073551, + 0.010668238624930382, + 0.010174321942031384, + 0.006587727461010218, + 0.007253784220665693, + 0.007809839211404324, + 0.009609386324882507, + 0.012555197812616825, + 0.018122751265764236, + 0.015098661184310913, + 0.007522932719439268, + 0.007737838663160801, + 0.01829325407743454, + 0.01161794364452362, + 0.010721887461841106, + 0.006665866356343031, + 0.012369057163596153, + 0.01886736787855625, + 0.016960129141807556, + 0.014149665832519531, + 0.018210595473647118, + 0.013617260381579399, + 0.00856790505349636, + 0.014947865158319473, + 0.010098189115524292, + 0.15487581491470337, + 0.018597163259983063, + 0.011093955487012863, + 0.16462300717830658, + 0.008309751749038696, + 0.011057881638407707, + 0.002639898331835866, + 0.002735947957262397, + 0.00479059386998415, + 0.0024746297858655453, + 0.0037830714136362076, + 0.005931640509516001, + 0.0033684996888041496, + 0.003675114829093218, + 0.003508966416120529, + 0.0042541432194411755, + 0.005712378304451704, + 0.003899132600054145, + 0.004424029029905796, + 0.0048897662200033665, + 0.0024680125061422586, + 0.004801004193723202, + 0.005455133970826864, + 0.00684709195047617, + 0.008352541364729404, + 0.006202945951372385, + 0.008437869139015675, + 0.011171507649123669, + 0.0064211818389594555, + 0.0072293574921786785, + 0.005224504042416811, + 0.005204170010983944, + 0.0017335849115625024, + 0.0011340952478349209, + 0.0013579040532931685, + 0.0014352691359817982, + 0.0016164317494258285, + 0.0020114947110414505, + 0.005349696613848209, + 0.017627520486712456, + 0.006534657906740904, + 0.008378813974559307, + 0.005811851471662521, + 0.004535700660198927, + 0.005891458597034216, + 0.015144657343626022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0019375376868993044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369060903787613 + ], + "opposite_logprob_tokens": [ + -3.4086906909942627 + ], + "logprob_reward": -0.03369060903787613, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369060903787613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20528027415275574, + 0.09509994834661484, + 0.11339870095252991, + 0.0910080298781395, + 0.1680205911397934, + 0.051184650510549545, + 0.03223148733377457, + 0.0759032666683197, + 0.04037010297179222, + 0.12750296294689178 + ], + "credit_map": [ + 0.021831722930073738, + 0.02543364092707634, + 0.028233669698238373, + 0.025253918021917343, + 0.019953669980168343, + 0.024747084826231003, + 0.03424607589840889, + 0.025580482557415962, + 0.011687771417200565, + 0.007707890588790178, + 0.015491731464862823, + 0.0073540350422263145, + 0.007543583400547504, + 0.014526008628308773, + 0.015965700149536133, + 0.01482322346419096, + 0.013664713129401207, + 0.016979649662971497, + 0.016147134825587273, + 0.008518392220139503, + 0.008559709414839745, + 0.02196677401661873, + 0.014231830835342407, + 0.013330493122339249, + 0.007336279843002558, + 0.006057439371943474, + 0.015889223664999008, + 0.01457762997597456, + 0.011333353817462921, + 0.014069583266973495, + 0.011703019961714745, + 0.010041499510407448, + 0.0074514769949018955, + 0.005307097919285297, + 0.06099793314933777, + 0.006949327420443296, + 0.009608877822756767, + 0.06366005539894104, + 0.006746023893356323, + 0.007299800869077444, + 0.004207538906484842, + 0.0047078519128263, + 0.0068749841302633286, + 0.007445249706506729, + 0.00955531932413578, + 0.007825924083590508, + 0.005234009586274624, + 0.00533377006649971, + 0.002767116529867053, + 0.0031117310281842947, + 0.0049364748410880566, + 0.003991628997027874, + 0.005011586472392082, + 0.0038813594728708267, + 0.004015547689050436, + 0.004516042303293943, + 0.006149179767817259, + 0.008906306698918343, + 0.009780227206647396, + 0.009261620230972767, + 0.014005853794515133, + 0.01308361068367958, + 0.007418392226099968, + 0.007298076990991831, + 0.009788753464818, + 0.007428219076246023, + 0.003358676331117749, + 0.002279447391629219, + 0.00178376620169729, + 0.002862086519598961, + 0.004316076636314392, + 0.00855307374149561, + 0.009858591482043266, + 0.032484013587236404, + 0.011388474144041538, + 0.012149189598858356, + 0.011343300342559814, + 0.009731732308864594, + 0.01071947906166315, + 0.029828179627656937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03369060903787613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633342333137989 + ], + "opposite_logprob_tokens": [ + -3.6513333320617676 + ], + "logprob_reward": -0.02633342333137989, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633342333137989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1465827226638794, + 0.08360379934310913, + 0.04883623868227005, + 0.05680897831916809, + 0.15675446391105652, + 0.04850945249199867, + 0.07829100638628006, + 0.13514462113380432, + 0.02985401451587677, + 0.21561475098133087 + ], + "credit_map": [ + 0.01352330669760704, + 0.018215026706457138, + 0.015411444939672947, + 0.015487282536923885, + 0.014335328713059425, + 0.022905386984348297, + 0.027311934158205986, + 0.01939300075173378, + 0.008842374198138714, + 0.00968214776366949, + 0.011480944231152534, + 0.015320357866585255, + 0.010192746296525002, + 0.00766837177798152, + 0.011317801661789417, + 0.009099045768380165, + 0.004420930054038763, + 0.0048128836788237095, + 0.007348985411226749, + 0.0071058934554457664, + 0.004792795516550541, + 0.006935483310371637, + 0.007499258033931255, + 0.005920005030930042, + 0.004697342403233051, + 0.0066617922857403755, + 0.008480548858642578, + 0.00640529440715909, + 0.0057989563792943954, + 0.01066624466329813, + 0.0072273071855306625, + 0.006871482357382774, + 0.006969565059989691, + 0.006854032166302204, + 0.04720107838511467, + 0.013607076369225979, + 0.008910675533115864, + 0.049634985625743866, + 0.016881756484508514, + 0.006695284508168697, + 0.004565634299069643, + 0.003883849363774061, + 0.0061301495879888535, + 0.005528062582015991, + 0.007516239769756794, + 0.007613955065608025, + 0.008232622407376766, + 0.005038934294134378, + 0.008253836072981358, + 0.011165379546582699, + 0.010985692031681538, + 0.008724691346287727, + 0.011314531788229942, + 0.010864367708563805, + 0.007740706205368042, + 0.009241794236004353, + 0.013788338750600815, + 0.027742872014641762, + 0.01588660664856434, + 0.013841720297932625, + 0.016835201531648636, + 0.024099843576550484, + 0.011745449155569077, + 0.011204566806554794, + 0.0045364024117589, + 0.007687605917453766, + 0.002963520120829344, + 0.0027034454979002476, + 0.00279635121114552, + 0.002518804045394063, + 0.002942899474874139, + 0.0037049828097224236, + 0.015005015768110752, + 0.030965732410550117, + 0.035753197968006134, + 0.027402160689234734, + 0.024458572268486023, + 0.02134178765118122, + 0.021091433241963387, + 0.03959682583808899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.02633342333137989 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic aerial drone shot slowly and steadily follows an eagle as it glides leisurely through the air. The eagle's massive wings are fully extended, each dark brown feather gleaming metallically in the golden morning sunlight, its edges sharply defined by the light. Its piercing yellow eyes gaze intently at the landscape below, its curved beak a look of resolve and power. Beneath it, a vast expanse of mountains stretches, its ridgeline undulating like the back of a giant dragon. Deep green coniferous forests cover most of the slopes, punctuated by exposed gray rock cliffs. A light morning mist lingers in the valley, adding a touch of mystery to the magnificent landscape. The eagle barely flaps its wings, skillfully riding the rising thermals in the valley, adjusting its flight with only subtle tilts of its body and the gentle waving of its tail feathers. Its flight is smooth and effortless. The sunlight shines down at a low angle, not only outlining a circle of bright outline light on the eagle's back, but also casting long, soft blue shadows on the back of the mountains. The whole picture presents an atmosphere of grandeur, tranquility and primitive vitality, full of awe of the power and freedom of nature.", + "seed": 32196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000097/group_6/index_1254_seed_32196_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03036008244380355, + "vlm_logprob": -0.03036008244380355, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03036008244380355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013464707881212234 + ], + "opposite_logprob_tokens": [ + -4.38846492767334 + ], + "logprob_reward": -0.013464707881212234, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013464707881212234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18237943947315216, + 0.17708908021450043, + 0.08521851152181625, + 0.031477756798267365, + 0.14434309303760529, + 0.14955122768878937, + 0.08429329842329025, + 0.031584884971380234, + 0.08390826731920242, + 0.030154531821608543 + ], + "credit_map": [ + 0.012495824135839939, + 0.012615050189197063, + 0.025435104966163635, + 0.0327620767056942, + 0.014466268010437489, + 0.030985865741968155, + 0.02568652480840683, + 0.02793271653354168, + 0.012519891373813152, + 0.015019419603049755, + 0.026265809312462807, + 0.02242765575647354, + 0.017289312556385994, + 0.03240099921822548, + 0.024128282442688942, + 0.027037687599658966, + 0.007244205567985773, + 0.006946745794266462, + 0.012969030998647213, + 0.009829753078520298, + 0.006708160508424044, + 0.009662730619311333, + 0.016012437641620636, + 0.015845436602830887, + 0.0030086699407547712, + 0.0026193943340331316, + 0.0048430017195641994, + 0.004261319525539875, + 0.0027295961044728756, + 0.004152401816099882, + 0.003904415760189295, + 0.005958953872323036, + 0.01633460633456707, + 0.007701585069298744, + 0.053949739784002304, + 0.01831022836267948, + 0.00546991964802146, + 0.007255487143993378, + 0.011514351703226566, + 0.02380715124309063, + 0.007309186737984419, + 0.006693193688988686, + 0.055800262838602066, + 0.012624501250684261, + 0.004751869011670351, + 0.005608813837170601, + 0.008163419552147388, + 0.0485999695956707, + 0.008822835981845856, + 0.01029188185930252, + 0.01481167133897543, + 0.010489440523087978, + 0.007218393497169018, + 0.009113384410738945, + 0.009401426650583744, + 0.014144256711006165, + 0.0037496977020055056, + 0.0032015761826187372, + 0.005938855465501547, + 0.004848796874284744, + 0.001956870313733816, + 0.0029327066149562597, + 0.003021926386281848, + 0.005934450775384903, + 0.00954297836869955, + 0.0023517680820077658, + 0.037226226180791855, + 0.021122844889760017, + 0.002277475781738758, + 0.002975714858621359, + 0.0030078303534537554, + 0.005403423681855202, + 0.0038039141800254583, + 0.00484021520242095, + 0.003944012802094221, + 0.003076280700042844, + 0.004756372421979904, + 0.0034898356534540653, + 0.002620606916025281, + 0.003623290453106165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.013464707881212234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901675999164581 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901675999164581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901675999164581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07513253390789032, + 0.11995409429073334, + 0.07297048717737198, + 0.02615939825773239, + 0.2537582218647003, + 0.04213926941156387, + 0.05539170280098915, + 0.05465491861104965, + 0.243317648768425, + 0.0565217100083828 + ], + "credit_map": [ + 0.007008040323853493, + 0.007009346969425678, + 0.011857625097036362, + 0.015919944271445274, + 0.008455806411802769, + 0.00786600261926651, + 0.008512001484632492, + 0.008503763005137444, + 0.01175934448838234, + 0.010381702333688736, + 0.014622555114328861, + 0.015447319485247135, + 0.012680871412158012, + 0.026405349373817444, + 0.015598653815686703, + 0.013058291748166084, + 0.0054131969809532166, + 0.0043503581546247005, + 0.012259794399142265, + 0.012342216446995735, + 0.006538082379847765, + 0.00704509112983942, + 0.00930606760084629, + 0.01571567729115486, + 0.002380054211243987, + 0.0027951307129114866, + 0.004243038594722748, + 0.0037332859355956316, + 0.002101171761751175, + 0.0032262192107737064, + 0.003536038566380739, + 0.0041444553062319756, + 0.01211792416870594, + 0.006561955437064171, + 0.10435743629932404, + 0.025649428367614746, + 0.004356136545538902, + 0.00938543863594532, + 0.008927015587687492, + 0.08240288496017456, + 0.0025316947139799595, + 0.0024878319818526506, + 0.01105932891368866, + 0.005616040900349617, + 0.0024444281589239836, + 0.0021548622753471136, + 0.006701253820210695, + 0.00914382841438055, + 0.00551407877355814, + 0.008179908618330956, + 0.008720935322344303, + 0.0075546507723629475, + 0.005584536120295525, + 0.005645620636641979, + 0.005600241012871265, + 0.008591728284955025, + 0.007551217917352915, + 0.006382503546774387, + 0.011310048401355743, + 0.007121301721781492, + 0.0034133652225136757, + 0.004318853374570608, + 0.004822302609682083, + 0.00973532348871231, + 0.024453671649098396, + 0.00604969123378396, + 0.10900698602199554, + 0.06786341220140457, + 0.006254279054701328, + 0.008393531665205956, + 0.007857606746256351, + 0.013438454829156399, + 0.008061750791966915, + 0.007475234568119049, + 0.008697886019945145, + 0.006967770867049694, + 0.005754954181611538, + 0.006664811633527279, + 0.005547394976019859, + 0.007351902313530445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.07901675999164581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031898352317512035 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.0031898352317512035, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031898352317512035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10993517935276031, + 0.08640632778406143, + 0.05179823189973831, + 0.020305003970861435, + 0.29523396492004395, + 0.13468068838119507, + 0.040169745683670044, + 0.05149742588400841, + 0.15510886907577515, + 0.05486459657549858 + ], + "credit_map": [ + 0.008551469072699547, + 0.009137454442679882, + 0.014431658200919628, + 0.018676478415727615, + 0.009964189492166042, + 0.01589587703347206, + 0.01575719565153122, + 0.017520863562822342, + 0.008807322010397911, + 0.008469413965940475, + 0.015736233443021774, + 0.0108147282153368, + 0.0051520708948373795, + 0.008607317693531513, + 0.011319954879581928, + 0.0174992848187685, + 0.004215060267597437, + 0.00591650977730751, + 0.00829659029841423, + 0.007458893116563559, + 0.0037397658452391624, + 0.00714311096817255, + 0.007562436629086733, + 0.007465867325663567, + 0.0015351326437667012, + 0.0019416940631344914, + 0.0031788854394108057, + 0.003140002256259322, + 0.0018296480411663651, + 0.002184248762205243, + 0.0033520592842251062, + 0.0031433335971087217, + 0.011239309795200825, + 0.006824520882219076, + 0.12072624266147614, + 0.029235705733299255, + 0.004731279332190752, + 0.009236527606844902, + 0.009865837171673775, + 0.10337453335523605, + 0.006581855937838554, + 0.005261717829853296, + 0.051630694419145584, + 0.009479077532887459, + 0.002986541483551264, + 0.00438856054097414, + 0.006983789149671793, + 0.047368451952934265, + 0.0030684969387948513, + 0.005034956149756908, + 0.007408224977552891, + 0.007771891541779041, + 0.0046004788018763065, + 0.0027085072360932827, + 0.004178660921752453, + 0.005398530978709459, + 0.004720382858067751, + 0.0047888257540762424, + 0.010500104166567326, + 0.008337624371051788, + 0.00447068689391017, + 0.003244641702622175, + 0.007173019461333752, + 0.008262140676379204, + 0.015755610540509224, + 0.005386142060160637, + 0.06499749422073364, + 0.04100155830383301, + 0.006953198928385973, + 0.0066438294015824795, + 0.005838470533490181, + 0.008532566018402576, + 0.006158431526273489, + 0.007196205668151379, + 0.007688140030950308, + 0.008299490436911583, + 0.0075658331625163555, + 0.005129864905029535, + 0.006308882962912321, + 0.006517750211060047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.0031898352317512035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340622454881668 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.026340622454881668, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340622454881668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16689273715019226, + 0.057873643934726715, + 0.10079577565193176, + 0.04365842416882515, + 0.05453154444694519, + 0.2869073152542114, + 0.07225029915571213, + 0.08481325954198837, + 0.05441140756011009, + 0.07786551862955093 + ], + "credit_map": [ + 0.016248691827058792, + 0.013934031128883362, + 0.021187664940953255, + 0.025648320093750954, + 0.0116384606808424, + 0.026787925511598587, + 0.02507709339261055, + 0.026370547711849213, + 0.005212495103478432, + 0.00707888463512063, + 0.008537322282791138, + 0.007483019959181547, + 0.00427427189424634, + 0.007547301705926657, + 0.007522077765315771, + 0.010218273848295212, + 0.006994565483182669, + 0.008963542059063911, + 0.01649440824985504, + 0.01413724571466446, + 0.007426061667501926, + 0.01682364195585251, + 0.014130436815321445, + 0.015825871378183365, + 0.0028419445734471083, + 0.0035659654531627893, + 0.00704246386885643, + 0.007490553893148899, + 0.003636422334238887, + 0.005747589748352766, + 0.006590195000171661, + 0.0067432899959385395, + 0.003658139845356345, + 0.0027795073110610247, + 0.01634710468351841, + 0.004983808845281601, + 0.0016326295444741845, + 0.0028104318771511316, + 0.0035182125866413116, + 0.01880171336233616, + 0.013760906644165516, + 0.0113285593688488, + 0.11668451130390167, + 0.021341951563954353, + 0.003283914178609848, + 0.00739275012165308, + 0.0096271438524127, + 0.10348758101463318, + 0.004660591017454863, + 0.00876402948051691, + 0.015428558923304081, + 0.01321170199662447, + 0.006838376633822918, + 0.0049032182432711124, + 0.0068134586326777935, + 0.011630366556346416, + 0.010528907179832458, + 0.008421122096478939, + 0.018421409651637077, + 0.011986755765974522, + 0.005012611858546734, + 0.005783262196928263, + 0.010186027735471725, + 0.014473162591457367, + 0.0026848765555769205, + 0.0029281696770340204, + 0.018416622653603554, + 0.016696659848093987, + 0.0030991127714514732, + 0.002719172043725848, + 0.004414488095790148, + 0.0034523068461567163, + 0.007586083374917507, + 0.0074863494373857975, + 0.01322227157652378, + 0.014688814990222454, + 0.009027963504195213, + 0.008696252480149269, + 0.008045818656682968, + 0.009111961349844933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.026340622454881668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle with fully extended dark brown wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's flight smooth and effortless with barely any wing flapping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast mountain landscape with deep green coniferous forests and gray rock cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a light morning mist lingering in the valley below the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature low-angle sunlight casting long soft blue shadows on the mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788486659526825 + ], + "opposite_logprob_tokens": [ + -3.5297884941101074 + ], + "logprob_reward": -0.029788486659526825, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788486659526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034012239426374435, + 0.03135836496949196, + 0.03757611662149429, + 0.02434028498828411, + 0.6178115010261536, + 0.0870242491364479, + 0.011288993991911411, + 0.03396277874708176, + 0.07995865494012833, + 0.042666852474212646 + ], + "credit_map": [ + 0.0025621100794523954, + 0.003463025903329253, + 0.004706154577434063, + 0.00507800979539752, + 0.002444795100018382, + 0.005082328338176012, + 0.004878196865320206, + 0.00579761853441596, + 0.0021283524110913277, + 0.0024611251428723335, + 0.0042080688290297985, + 0.004155424423515797, + 0.002812393009662628, + 0.005222531966865063, + 0.004611027892678976, + 0.0057594384998083115, + 0.003460274077951908, + 0.003624074161052704, + 0.0060798851773142815, + 0.003686404787003994, + 0.002478252863511443, + 0.003780034836381674, + 0.00445968983694911, + 0.01000750157982111, + 0.0031325225718319416, + 0.002680143341422081, + 0.003865325590595603, + 0.003104711417108774, + 0.002075849799439311, + 0.002699402393773198, + 0.0028351955115795135, + 0.003947136458009481, + 0.025961223989725113, + 0.014721667394042015, + 0.25574538111686707, + 0.06287050247192383, + 0.009890171699225903, + 0.01855562813580036, + 0.02037600427865982, + 0.2096908986568451, + 0.00425185589119792, + 0.003363034687936306, + 0.03460010513663292, + 0.006428869906812906, + 0.0012330535100772977, + 0.002599539468064904, + 0.003162802429869771, + 0.031384989619255066, + 0.0009308466105721891, + 0.001001963741146028, + 0.0020859455689787865, + 0.0024126723874360323, + 0.0008934300276450813, + 0.0011816512560471892, + 0.0011887401342391968, + 0.00159374438226223, + 0.003191194962710142, + 0.0031603791285306215, + 0.005371742881834507, + 0.005068605300039053, + 0.0023727768566459417, + 0.003646896220743656, + 0.00559609942138195, + 0.005555082578212023, + 0.008951116353273392, + 0.003936694469302893, + 0.032685380429029465, + 0.019871754571795464, + 0.0034634897019714117, + 0.0032594737131148577, + 0.0030155188869684935, + 0.004775217734277248, + 0.0058510201051831245, + 0.006658267229795456, + 0.006802922114729881, + 0.003932978492230177, + 0.005907873157411814, + 0.0036562555469572544, + 0.0040349499322474, + 0.005822585895657539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.029788486659526825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_tensor([1422, 1422])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_tensor([1422, 1422])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5262ba13bee6c47fc243d62e3503ba4d2f8b8bf3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_tensor([1422, 1422])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tilted, extreme close-up captures a solid bat emblem, forged from cold, subtly scratched black metal, its outline imbued with tension. Against the backdrop of a blurred, rain-soaked, Gothic cityscape, flickering crimson and electric blue neon lights cast a shadow on the emblem's surface. The emblem begins to dissolve precariously from its sharp edges, its smooth metallic texture transforming into a thick, oily, inky liquid that slowly drips downward, forming plump droplets. Rendered with the exquisite clarity of ray tracing, each falling droplet resembles a miniature convex mirror, accurately reflecting and distorting the surrounding shimmering neon signs and cold city lights. The slow, eerie interplay of liquid flow and reflection creates a surreal visual spectacle set in a dark, ominous, and technologically futuristic atmosphere.", + "seed": 4519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_1422_seed_4519_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23685167715884745, + "vlm_logprob": -0.23685167715884745, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.23685167715884745, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.055582880973816 + ], + "opposite_logprob_tokens": [ + -0.43058285117149353 + ], + "logprob_reward": -1.055582880973816, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.055582880973816, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16214075684547424, + 0.14341367781162262, + 0.06448331475257874, + 0.16629351675510406, + 0.06900092959403992, + 0.040008530020713806, + 0.07307194918394089, + 0.06820179522037506, + 0.1394725888967514, + 0.07391292601823807 + ], + "credit_map": [ + 0.018333710730075836, + 0.023211563006043434, + 0.020551258698105812, + 0.016929250210523605, + 0.020225346088409424, + 0.019097639247775078, + 0.018671460449695587, + 0.025120532140135765, + 0.014692007564008236, + 0.023592177778482437, + 0.01669846661388874, + 0.013179609552025795, + 0.023883895948529243, + 0.01086344476789236, + 0.020662691444158554, + 0.019841374829411507, + 0.00699614966288209, + 0.007419118192046881, + 0.011282622814178467, + 0.007897896692156792, + 0.008509178645908833, + 0.006537817418575287, + 0.0070986514911055565, + 0.008741883561015129, + 0.02105853147804737, + 0.024003444239497185, + 0.02050905115902424, + 0.012571049854159355, + 0.030968770384788513, + 0.015547825023531914, + 0.02080959640443325, + 0.020825259387493134, + 0.00960946548730135, + 0.00761942844837904, + 0.0073386263102293015, + 0.00888026226311922, + 0.009959006682038307, + 0.008801224641501904, + 0.007109133992344141, + 0.009683785028755665, + 0.006739517208188772, + 0.005127900745719671, + 0.002891858574002981, + 0.006387947592884302, + 0.004715498071163893, + 0.003048230893909931, + 0.004928433336317539, + 0.006169139873236418, + 0.006665404886007309, + 0.010621662251651287, + 0.008932297118008137, + 0.011322438716888428, + 0.007559095975011587, + 0.007751120254397392, + 0.010297957807779312, + 0.009921969845890999, + 0.00871709268540144, + 0.009616713970899582, + 0.0057028718292713165, + 0.0043934364803135395, + 0.007271131966263056, + 0.01087992824614048, + 0.010149593465030193, + 0.011471026577055454, + 0.011788463220000267, + 0.010113054886460304, + 0.011714358814060688, + 0.01677640713751316, + 0.013386559672653675, + 0.02819957211613655, + 0.027437357231974602, + 0.020056819543242455, + 0.006501168943941593, + 0.0075449394062161446, + 0.007492223754525185, + 0.0075127677991986275, + 0.008074133656919003, + 0.010444270446896553, + 0.014972927048802376, + 0.011370491236448288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -1.055582880973816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038159787654877 + ], + "opposite_logprob_tokens": [ + -2.35038161277771 + ], + "logprob_reward": -0.10038159787654877, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038159787654877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08915886282920837, + 0.15696896612644196, + 0.053676899522542953, + 0.13405421376228333, + 0.09632183611392975, + 0.028738966211676598, + 0.03861188143491745, + 0.051990315318107605, + 0.1043398380279541, + 0.24613815546035767 + ], + "credit_map": [ + 0.011617022566497326, + 0.012355642393231392, + 0.00936088990420103, + 0.009128128178417683, + 0.01559445820748806, + 0.006360398605465889, + 0.009567863307893276, + 0.01517445407807827, + 0.019168181344866753, + 0.028845731168985367, + 0.02097819373011589, + 0.015083092264831066, + 0.02016231045126915, + 0.00949107576161623, + 0.017063738778233528, + 0.02617664448916912, + 0.007861824706196785, + 0.0064254505559802055, + 0.006581670604646206, + 0.0039141494780778885, + 0.007273098919540644, + 0.00587238697335124, + 0.008029518648982048, + 0.007718800101429224, + 0.019402360543608665, + 0.01824990101158619, + 0.018014920875430107, + 0.012095870450139046, + 0.02135990932583809, + 0.011521777138113976, + 0.01683623529970646, + 0.016573240980505943, + 0.013382510282099247, + 0.011207657866179943, + 0.007734921760857105, + 0.01137466449290514, + 0.013963533565402031, + 0.013913838192820549, + 0.010470852255821228, + 0.014273862354457378, + 0.003465590300038457, + 0.0038828018587082624, + 0.00308958999812603, + 0.00428813649341464, + 0.0031796128023415804, + 0.0029861340299248695, + 0.003203809028491378, + 0.004643293097615242, + 0.0032274441327899694, + 0.0059248958714306355, + 0.004531182814389467, + 0.00599371362477541, + 0.004016523249447346, + 0.0032768240198493004, + 0.00560292461887002, + 0.006038373801857233, + 0.005104545038193464, + 0.008721091784536839, + 0.003726392751559615, + 0.004534356761723757, + 0.007231933530420065, + 0.006933669559657574, + 0.00870424509048462, + 0.007034081965684891, + 0.010476235300302505, + 0.008828872814774513, + 0.0075677623972296715, + 0.011729325167834759, + 0.00985557772219181, + 0.01788884401321411, + 0.02411087229847908, + 0.013882352039217949, + 0.024341126903891563, + 0.018605085089802742, + 0.020641231909394264, + 0.023503471165895462, + 0.023056138306856155, + 0.041503287851810455, + 0.05009393393993378, + 0.044393885880708694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.10038159787654877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028206822462379932 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028206822462379932, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028206822462379932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2120566964149475, + 0.052394647151231766, + 0.053569868206977844, + 0.15063761174678802, + 0.05382232367992401, + 0.03479953110218048, + 0.029688239097595215, + 0.03591899946331978, + 0.02733243629336357, + 0.3497796654701233 + ], + "credit_map": [ + 0.019428441300988197, + 0.037805911153554916, + 0.014575546607375145, + 0.01744098775088787, + 0.04442133009433746, + 0.028097808361053467, + 0.01533473189920187, + 0.03495193272829056, + 0.006897482089698315, + 0.008366799913346767, + 0.0048035248182713985, + 0.004618436563760042, + 0.007241038139909506, + 0.0056704566814005375, + 0.0072138044051826, + 0.007583103142678738, + 0.0071838428266346455, + 0.00850357860326767, + 0.004991909489035606, + 0.004451126325875521, + 0.00786304846405983, + 0.007295973598957062, + 0.0046080127358436584, + 0.008672374300658703, + 0.020126022398471832, + 0.021841304376721382, + 0.017353776842355728, + 0.012661182321608067, + 0.026556475088000298, + 0.014178634621202946, + 0.018396729603409767, + 0.019523490220308304, + 0.0059575573541224, + 0.007579307071864605, + 0.0057294657453894615, + 0.0063899788074195385, + 0.008130517788231373, + 0.0067055560648441315, + 0.005385442171245813, + 0.007944491691887379, + 0.005336201749742031, + 0.004979254677891731, + 0.0038143813144415617, + 0.003720255568623543, + 0.004540808033198118, + 0.0036310169380158186, + 0.0037033024709671736, + 0.0050743091851472855, + 0.0024131215177476406, + 0.004370954819023609, + 0.002467117039486766, + 0.004202128853648901, + 0.0033909096382558346, + 0.0034219908993691206, + 0.004571957513689995, + 0.0048500546254217625, + 0.0037835713010281324, + 0.0042359475046396255, + 0.0035796242300421, + 0.0027653914876282215, + 0.003913084976375103, + 0.0044722226448357105, + 0.006486508529633284, + 0.0066826483234763145, + 0.002249035518616438, + 0.0029077695216983557, + 0.0023997011594474316, + 0.0038080483209341764, + 0.0027589050587266684, + 0.004323829896748066, + 0.0037742010317742825, + 0.005110946483910084, + 0.025085419416427612, + 0.023532461374998093, + 0.034376706928014755, + 0.0326593853533268, + 0.03218958526849747, + 0.060829706490039825, + 0.07382623851299286, + 0.06728015840053558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0028206822462379932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198661444708705 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198661444708705, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198661444708705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15136490762233734, + 0.1703723520040512, + 0.05493716150522232, + 0.14607110619544983, + 0.047130052000284195, + 0.02982737496495247, + 0.02856719121336937, + 0.041170306503772736, + 0.06240065023303032, + 0.2681589126586914 + ], + "credit_map": [ + 0.021359160542488098, + 0.020214898511767387, + 0.020419146865606308, + 0.012156892567873001, + 0.02237755060195923, + 0.0178447924554348, + 0.012447415851056576, + 0.024545036256313324, + 0.021686147898435593, + 0.017377866432070732, + 0.024024037644267082, + 0.016786428168416023, + 0.021406520158052444, + 0.013803255744278431, + 0.02895190194249153, + 0.02633618749678135, + 0.007154191378504038, + 0.0053556375205516815, + 0.009559174999594688, + 0.006731898058205843, + 0.006595077458769083, + 0.004733601585030556, + 0.007727928925305605, + 0.007079652044922113, + 0.020535359159111977, + 0.020942796021699905, + 0.01857110857963562, + 0.012435299344360828, + 0.02536037750542164, + 0.01230473443865776, + 0.01786247082054615, + 0.018058959394693375, + 0.004928878974169493, + 0.00769902765750885, + 0.0037674217019230127, + 0.004774568136781454, + 0.008830033242702484, + 0.004600405693054199, + 0.005749656353145838, + 0.006780058611184359, + 0.003962006885558367, + 0.004595611710101366, + 0.0022584018297493458, + 0.0047868178226053715, + 0.00352227920666337, + 0.0027002072893083096, + 0.0033374750055372715, + 0.004664575215429068, + 0.0021449641790241003, + 0.003182861255481839, + 0.0020833490416407585, + 0.0038959768135100603, + 0.0040204175747931, + 0.0031035856809467077, + 0.005412265658378601, + 0.004723771940916777, + 0.004067940171808004, + 0.00398681778460741, + 0.003655391512438655, + 0.0031162246596068144, + 0.004450453445315361, + 0.0067065744660794735, + 0.008156130090355873, + 0.007030776236206293, + 0.004031018353998661, + 0.004374700598418713, + 0.006309844553470612, + 0.0051727863028645515, + 0.0055306958965957165, + 0.013927021995186806, + 0.011439885012805462, + 0.011614696122705936, + 0.020822718739509583, + 0.02052912674844265, + 0.02568703331053257, + 0.025865089148283005, + 0.026738720014691353, + 0.04179973155260086, + 0.05933458358049393, + 0.0473819375038147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.002198661444708705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274563252925873 + ], + "opposite_logprob_tokens": [ + -3.7732746601104736 + ], + "logprob_reward": -0.023274563252925873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274563252925873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08628229051828384, + 0.09034707397222519, + 0.02483457326889038, + 0.05082804337143898, + 0.03754017874598503, + 0.05849605053663254, + 0.02119619958102703, + 0.051748745143413544, + 0.2035520225763321, + 0.37517476081848145 + ], + "credit_map": [ + 0.009529243223369122, + 0.014961844310164452, + 0.007834305055439472, + 0.006841962691396475, + 0.01464314665645361, + 0.012266380712389946, + 0.005704863928258419, + 0.014500552788376808, + 0.010556463152170181, + 0.007770075462758541, + 0.013577507808804512, + 0.011771216057240963, + 0.00890823733061552, + 0.007002581842243671, + 0.01738496497273445, + 0.013376026414334774, + 0.002720214892178774, + 0.0022935643792152405, + 0.004682701546698809, + 0.004449031315743923, + 0.002271799836307764, + 0.002366319065913558, + 0.0023682648316025734, + 0.0036826760042458773, + 0.00683391559869051, + 0.007067461963742971, + 0.008432838134467602, + 0.004476781003177166, + 0.008102307096123695, + 0.003643334610387683, + 0.0062357052229344845, + 0.006035701371729374, + 0.004131869878619909, + 0.0056985486298799515, + 0.003966524265706539, + 0.004184770863503218, + 0.00629828218370676, + 0.003232828574255109, + 0.004696310497820377, + 0.005331041757017374, + 0.008550460450351238, + 0.010392055846750736, + 0.0046062171459198, + 0.008226507343351841, + 0.0051123956218361855, + 0.004753405228257179, + 0.006742783822119236, + 0.010112222284078598, + 0.0018812764901667833, + 0.0027654562145471573, + 0.0015088163781911135, + 0.002441723830997944, + 0.0030811342876404524, + 0.00198911945335567, + 0.004173302091658115, + 0.003355371532961726, + 0.004265566356480122, + 0.007724551018327475, + 0.0036125448532402515, + 0.005038673523813486, + 0.00821524765342474, + 0.007644703146070242, + 0.008584138937294483, + 0.006663323845714331, + 0.009662838652729988, + 0.00960586778819561, + 0.015228722244501114, + 0.02374933660030365, + 0.01641884073615074, + 0.04374954104423523, + 0.05081160366535187, + 0.034325286746025085, + 0.028103439137339592, + 0.028862886130809784, + 0.0351630374789238, + 0.03697394207119942, + 0.03780417516827583, + 0.0606628879904747, + 0.07929877936840057, + 0.0683056116104126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.023274563252925873 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, extreme close-up captures a solid bat emblem, forged from cold, subtly scratched black metal, its outline imbued with tension. Against the backdrop of a blurred, rain-soaked, Gothic cityscape, flickering crimson and electric blue neon lights cast a shadow on the emblem's surface. The emblem begins to dissolve precariously from its sharp edges, its smooth metallic texture transforming into a thick, oily, inky liquid that slowly drips downward, forming plump droplets. Rendered with the exquisite clarity of ray tracing, each falling droplet resembles a miniature convex mirror, accurately reflecting and distorting the surrounding shimmering neon signs and cold city lights. The slow, eerie interplay of liquid flow and reflection creates a surreal visual spectacle set in a dark, ominous, and technologically futuristic atmosphere.", + "seed": 4520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_1422_seed_4520_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.035789645928889516, + "vlm_logprob": -0.035789645928889516, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.035789645928889516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09121304750442505 + ], + "opposite_logprob_tokens": [ + -2.4662129878997803 + ], + "logprob_reward": -0.09121304750442505, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09121304750442505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17377300560474396, + 0.08667618036270142, + 0.056638311594724655, + 0.04521600902080536, + 0.09500128775835037, + 0.11395789682865143, + 0.1858782023191452, + 0.07543780654668808, + 0.07085881382226944, + 0.09656243771314621 + ], + "credit_map": [ + 0.010687153786420822, + 0.01691521145403385, + 0.02219751849770546, + 0.023671040311455727, + 0.022495385259389877, + 0.03808532655239105, + 0.025601791217923164, + 0.01411957573145628, + 0.006338993553072214, + 0.01046713162213564, + 0.015889069065451622, + 0.006263852119445801, + 0.0216551274061203, + 0.008792501874268055, + 0.006638405378907919, + 0.010631103999912739, + 0.004267544485628605, + 0.006333081517368555, + 0.008563975803554058, + 0.006038461811840534, + 0.007206450216472149, + 0.005543089471757412, + 0.005044019315391779, + 0.013641690835356712, + 0.004235782194882631, + 0.005068422295153141, + 0.006941480096429586, + 0.005386249162256718, + 0.005807290319353342, + 0.005011883098632097, + 0.006294525694102049, + 0.006470379885286093, + 0.008602694608271122, + 0.01629793271422386, + 0.012623369693756104, + 0.012220951728522778, + 0.011539490893483162, + 0.012023979797959328, + 0.008293749764561653, + 0.013399120420217514, + 0.00630961125716567, + 0.00991389062255621, + 0.03757396712899208, + 0.006261896807700396, + 0.004268139600753784, + 0.008473952300846577, + 0.00436257291585207, + 0.03679386153817177, + 0.0179651640355587, + 0.018354102969169617, + 0.03151852637529373, + 0.01821260154247284, + 0.018893878906965256, + 0.024721456691622734, + 0.019840549677610397, + 0.03637193515896797, + 0.012385580688714981, + 0.010534665547311306, + 0.010414882563054562, + 0.010596396401524544, + 0.007656609173864126, + 0.011626209132373333, + 0.005807589739561081, + 0.006415874231606722, + 0.010229501873254776, + 0.0109917763620615, + 0.01150250993669033, + 0.008503112941980362, + 0.005646891891956329, + 0.010420531965792179, + 0.006248889025300741, + 0.007315601222217083, + 0.01489748153835535, + 0.007345382124185562, + 0.014790834859013557, + 0.01227337121963501, + 0.015016138553619385, + 0.009099596180021763, + 0.00995681993663311, + 0.013182812370359898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.09121304750442505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211989007890224 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014211989007890224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211989007890224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057873666286468506, + 0.09099225699901581, + 0.023438137024641037, + 0.02593865618109703, + 0.046549588441848755, + 0.056747689843177795, + 0.5707107782363892, + 0.04490383714437485, + 0.03992738947272301, + 0.042918045073747635 + ], + "credit_map": [ + 0.004527336452156305, + 0.00614756578579545, + 0.009255809709429741, + 0.010631851851940155, + 0.006476836744695902, + 0.010799719020724297, + 0.005035610869526863, + 0.004998937249183655, + 0.0061447094194591045, + 0.016576381400227547, + 0.015188551507890224, + 0.010998059064149857, + 0.011543374508619308, + 0.006890743970870972, + 0.0085448008030653, + 0.01510563213378191, + 0.0018372951308265328, + 0.001719996565952897, + 0.003621421754360199, + 0.003149249590933323, + 0.002271275268867612, + 0.0028134225867688656, + 0.0030590298119932413, + 0.004966446198523045, + 0.0030513585079461336, + 0.0022099576890468597, + 0.00518309511244297, + 0.0029806667007505894, + 0.003141380613669753, + 0.0027641933411359787, + 0.0031868715304881334, + 0.003421131521463394, + 0.005476657301187515, + 0.0062373001128435135, + 0.006677215453237295, + 0.0033650093246251345, + 0.004505753982812166, + 0.0070260511711239815, + 0.00469035841524601, + 0.008571244776248932, + 0.0054571619257330894, + 0.004951731767505407, + 0.013386493548750877, + 0.004908578936010599, + 0.004444091115146875, + 0.0053987326100468636, + 0.005138085223734379, + 0.013062810525298119, + 0.026107829064130783, + 0.02847159653902054, + 0.18717418611049652, + 0.03447530046105385, + 0.018629226833581924, + 0.0439772829413414, + 0.018601953983306885, + 0.21327340602874756, + 0.008052566088736057, + 0.008995507843792439, + 0.006546580232679844, + 0.004047963302582502, + 0.007104387506842613, + 0.004282732494175434, + 0.0027693170122802258, + 0.003104784758761525, + 0.004694243893027306, + 0.004388711880892515, + 0.006232875864952803, + 0.005877133924514055, + 0.004773983731865883, + 0.0038582507986575365, + 0.004766618367284536, + 0.005335570778697729, + 0.0049563986249268055, + 0.005980828311294317, + 0.006386109162122011, + 0.00767487334087491, + 0.006063289940357208, + 0.0038145617581903934, + 0.004073652438819408, + 0.003968331962823868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.014211989007890224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06697934120893478, + 0.05382022634148598, + 0.05353136360645294, + 0.03578934445977211, + 0.040262557566165924, + 0.04304277151823044, + 0.44592732191085815, + 0.07647540420293808, + 0.07046610862016678, + 0.113705575466156 + ], + "credit_map": [ + 0.006170119624584913, + 0.006395568605512381, + 0.010823249816894531, + 0.00950551126152277, + 0.009077035821974277, + 0.009954508394002914, + 0.007551291957497597, + 0.00750206271186471, + 0.004007560666650534, + 0.005107378587126732, + 0.01128346286714077, + 0.006944595370441675, + 0.010519789531826973, + 0.005973832216113806, + 0.005010212771594524, + 0.004973397124558687, + 0.004400867037475109, + 0.003607087302953005, + 0.008374660275876522, + 0.0049958904273808, + 0.006401978898793459, + 0.005307817831635475, + 0.005012401845306158, + 0.015430659055709839, + 0.004903445020318031, + 0.0032161478884518147, + 0.005218243692070246, + 0.0036974886897951365, + 0.004527391865849495, + 0.0032983762212097645, + 0.0034670524764806032, + 0.007461200468242168, + 0.005078726448118687, + 0.004513579420745373, + 0.005588473752140999, + 0.003813644405454397, + 0.0055076745338737965, + 0.004815028980374336, + 0.005954520311206579, + 0.004990911111235619, + 0.005270072259008884, + 0.009934053756296635, + 0.006288518197834492, + 0.00415946077555418, + 0.0034057763405144215, + 0.005211900919675827, + 0.0028904047794640064, + 0.0058825877495110035, + 0.025956138968467712, + 0.018192006275057793, + 0.12417519837617874, + 0.029992742463946342, + 0.027104170992970467, + 0.048097267746925354, + 0.02982497774064541, + 0.14258481562137604, + 0.01538144052028656, + 0.010649222880601883, + 0.012511903420090675, + 0.009131772443652153, + 0.006525969598442316, + 0.008297054097056389, + 0.006355518940836191, + 0.007622518576681614, + 0.011599698103964329, + 0.009216338396072388, + 0.011203935369849205, + 0.008570508100092411, + 0.009297139011323452, + 0.006793717388063669, + 0.009347745217382908, + 0.004437025170773268, + 0.018884724006056786, + 0.01693081110715866, + 0.018148956820368767, + 0.022774819284677505, + 0.013762115500867367, + 0.0073316264897584915, + 0.008021349087357521, + 0.007851170375943184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08184169977903366, + 0.08367925882339478, + 0.04910504072904587, + 0.03987869992852211, + 0.04718660190701485, + 0.12831319868564606, + 0.2845073938369751, + 0.08259671181440353, + 0.09299761056900024, + 0.10989373922348022 + ], + "credit_map": [ + 0.00648528290912509, + 0.00789487361907959, + 0.011826155707240105, + 0.00894861575216055, + 0.0110129090026021, + 0.011832083575427532, + 0.015077591873705387, + 0.008764190599322319, + 0.006113024428486824, + 0.012208880856633186, + 0.015915965661406517, + 0.009223992004990578, + 0.012256708927452564, + 0.006620876956731081, + 0.005040111020207405, + 0.01629970408976078, + 0.003016712376847863, + 0.002434077672660351, + 0.012279143556952477, + 0.0039000839460641146, + 0.0027233229484409094, + 0.0037167223636060953, + 0.004282984416931868, + 0.016751999035477638, + 0.004617690574377775, + 0.004098632372915745, + 0.005363024305552244, + 0.0048951865173876286, + 0.006348302587866783, + 0.004226316697895527, + 0.005651460960507393, + 0.004678086377680302, + 0.002954948227852583, + 0.005239206366240978, + 0.0069723292253911495, + 0.005314059555530548, + 0.00599692901596427, + 0.0051230513490736485, + 0.008107186295092106, + 0.0074788969941437244, + 0.010406640358269215, + 0.008302860893309116, + 0.03597278147935867, + 0.009253453463315964, + 0.009230982512235641, + 0.012233367189764977, + 0.00733185512945056, + 0.03558127582073212, + 0.019581163302063942, + 0.01717468537390232, + 0.07763651013374329, + 0.01992548070847988, + 0.016747567802667618, + 0.030424179509282112, + 0.015848232433199883, + 0.0871695950627327, + 0.017385726794600487, + 0.010123686864972115, + 0.012357442639768124, + 0.006959984079003334, + 0.013246883638203144, + 0.009798639453947544, + 0.006005541421473026, + 0.006718816235661507, + 0.008970667608082294, + 0.00569630041718483, + 0.015831299126148224, + 0.007656286936253309, + 0.00844663754105568, + 0.0180218406021595, + 0.009570755995810032, + 0.018803833052515984, + 0.01690501905977726, + 0.01117156445980072, + 0.017727836966514587, + 0.016131242737174034, + 0.012770866975188255, + 0.01550253201276064, + 0.011333150789141655, + 0.008351534605026245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0013360392767935991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06998991966247559 + ], + "opposite_logprob_tokens": [ + -2.6949899196624756 + ], + "logprob_reward": -0.06998991966247559, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998991966247559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12964344024658203, + 0.16383054852485657, + 0.019179977476596832, + 0.04818270727992058, + 0.026304876431822777, + 0.15277229249477386, + 0.10650500655174255, + 0.10958866775035858, + 0.1912267506122589, + 0.05276568606495857 + ], + "credit_map": [ + 0.01384303905069828, + 0.017035730183124542, + 0.014734601601958275, + 0.012569926679134369, + 0.02072148583829403, + 0.017635155469179153, + 0.01999025046825409, + 0.013113253749907017, + 0.013073156587779522, + 0.03882420063018799, + 0.01901988685131073, + 0.03122306615114212, + 0.024254832416772842, + 0.012184311635792255, + 0.012352968566119671, + 0.012898110784590244, + 0.0016649486497044563, + 0.0015959390439093113, + 0.003482345724478364, + 0.002272052923217416, + 0.002126153092831373, + 0.002110318746417761, + 0.0023375831078737974, + 0.003590635024011135, + 0.00445402879267931, + 0.004445715341717005, + 0.006493886932730675, + 0.006900668144226074, + 0.007269228342920542, + 0.005027732811868191, + 0.006686442997306585, + 0.006905003916472197, + 0.002802191535010934, + 0.004541997332125902, + 0.003727682400494814, + 0.0037993346340954304, + 0.003198242047801614, + 0.0035610035993158817, + 0.001976351486518979, + 0.002698072697967291, + 0.009509757161140442, + 0.008774891495704651, + 0.054775696247816086, + 0.006509855389595032, + 0.0064863357692956924, + 0.007938274182379246, + 0.0046045854687690735, + 0.05417289584875107, + 0.006116791162639856, + 0.007283625192940235, + 0.027495205402374268, + 0.006418565288186073, + 0.007500112056732178, + 0.012549248524010181, + 0.00905881728976965, + 0.030082643032073975, + 0.025197042152285576, + 0.014238666743040085, + 0.016171196475625038, + 0.010963404551148415, + 0.01207553967833519, + 0.013243124820291996, + 0.008065400645136833, + 0.009634300135076046, + 0.02563966065645218, + 0.020783811807632446, + 0.03029194474220276, + 0.025191061198711395, + 0.01645137555897236, + 0.026699457317590714, + 0.019646884873509407, + 0.026522567495703697, + 0.008220546878874302, + 0.004370649345219135, + 0.007715051062405109, + 0.009217500686645508, + 0.004681192338466644, + 0.006106675136834383, + 0.004599356092512608, + 0.00785471498966217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.06998991966247559 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, extreme close-up captures a solid bat emblem, forged from cold, subtly scratched black metal, its outline imbued with tension. Against the backdrop of a blurred, rain-soaked, Gothic cityscape, flickering crimson and electric blue neon lights cast a shadow on the emblem's surface. The emblem begins to dissolve precariously from its sharp edges, its smooth metallic texture transforming into a thick, oily, inky liquid that slowly drips downward, forming plump droplets. Rendered with the exquisite clarity of ray tracing, each falling droplet resembles a miniature convex mirror, accurately reflecting and distorting the surrounding shimmering neon signs and cold city lights. The slow, eerie interplay of liquid flow and reflection creates a surreal visual spectacle set in a dark, ominous, and technologically futuristic atmosphere.", + "seed": 4521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_1422_seed_4521_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08272063401527703, + "vlm_logprob": -0.08272063401527703, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.08272063401527703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35088491439819336 + ], + "opposite_logprob_tokens": [ + -1.2258849143981934 + ], + "logprob_reward": -0.35088491439819336, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35088491439819336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02350805513560772, + 0.8328248858451843, + 0.010345946997404099, + 0.018960926681756973, + 0.0112539604306221, + 0.013764218427240849, + 0.01285835262387991, + 0.029837239533662796, + 0.02738068625330925, + 0.019265709444880486 + ], + "credit_map": [ + 0.0024790745228528976, + 0.0029736619908362627, + 0.0033435754012316465, + 0.003266735700890422, + 0.003052858170121908, + 0.0023657481651753187, + 0.002898954087868333, + 0.003127449192106724, + 0.05765051767230034, + 0.01733999513089657, + 0.3213346600532532, + 0.029142899438738823, + 0.0165239404886961, + 0.3565625250339508, + 0.01608581468462944, + 0.0181846022605896, + 0.0011110916966572404, + 0.0009619049960747361, + 0.0018038512207567692, + 0.0011619169963523746, + 0.0010315091349184513, + 0.002072287490591407, + 0.0008345976821146905, + 0.0013687886530533433, + 0.0019304692978039384, + 0.0019460421754047275, + 0.0027234493754804134, + 0.002660389756783843, + 0.0021615088917315006, + 0.002646544249728322, + 0.002154999179765582, + 0.0027375249192118645, + 0.0014009318547323346, + 0.0013875040458515286, + 0.0014705510111525655, + 0.0010548941791057587, + 0.0008983865263871849, + 0.0011612384114414454, + 0.0019351517548784614, + 0.0019453028216958046, + 0.0008710419060662389, + 0.0011209436925128102, + 0.0023555310908704996, + 0.001641544746235013, + 0.001785491593182087, + 0.0021645333617925644, + 0.0023294040001928806, + 0.0014957290841266513, + 0.0009129107929766178, + 0.000981019577011466, + 0.0018218207405880094, + 0.0015288536669686437, + 0.0010819309391081333, + 0.0015323732513934374, + 0.0018118314910680056, + 0.00318761239759624, + 0.002474616514518857, + 0.0019486872479319572, + 0.0032806468661874533, + 0.007328297942876816, + 0.0027426271699368954, + 0.0027766874991357327, + 0.0053237611427903175, + 0.003961917012929916, + 0.0033551277592778206, + 0.002687524538487196, + 0.0034775445237755775, + 0.002993987873196602, + 0.0033451011404395103, + 0.003787560388445854, + 0.0045728241093456745, + 0.0031610182486474514, + 0.002912260126322508, + 0.0029965986032038927, + 0.0021746500860899687, + 0.0019844183698296547, + 0.002769692102447152, + 0.0027115177363157272, + 0.00205521984025836, + 0.001661352813243866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.35088491439819336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819844290614128 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819844290614128, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819844290614128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05966443568468094, + 0.6471307873725891, + 0.011678969487547874, + 0.043575845658779144, + 0.035709768533706665, + 0.020249642431735992, + 0.025422299280762672, + 0.03373631462454796, + 0.043901022523641586, + 0.07893090695142746 + ], + "credit_map": [ + 0.007444378454238176, + 0.007092248648405075, + 0.009394340217113495, + 0.008503374643623829, + 0.005928263999521732, + 0.007942440919578075, + 0.005674383137375116, + 0.007685008458793163, + 0.04489675164222717, + 0.013993426226079464, + 0.24847747385501862, + 0.02257489040493965, + 0.01343198586255312, + 0.2764533460140228, + 0.01277729868888855, + 0.014525599777698517, + 0.001310947467572987, + 0.0012178239412605762, + 0.0016103541711345315, + 0.001157425926066935, + 0.002251085825264454, + 0.0017790991114452481, + 0.0008843218092806637, + 0.001467911177314818, + 0.004110202193260193, + 0.004066548775881529, + 0.006455992814153433, + 0.005667926277965307, + 0.004936198238283396, + 0.006354128010571003, + 0.005487790331244469, + 0.006497058551758528, + 0.0031145773828029633, + 0.0024631309788674116, + 0.0059801992028951645, + 0.0032777017913758755, + 0.004030558280646801, + 0.004904391244053841, + 0.005416418891400099, + 0.006522789131850004, + 0.001554133603349328, + 0.002126380568370223, + 0.0024758006911724806, + 0.0030990999657660723, + 0.0025710021145641804, + 0.0029794909060001373, + 0.00353847979567945, + 0.0019052544375881553, + 0.0027826905716210604, + 0.0017455759225413203, + 0.0031338962726294994, + 0.002870881697162986, + 0.002047883812338114, + 0.002833431353792548, + 0.0034876097925007343, + 0.006520329974591732, + 0.0025116188917309046, + 0.002537665655836463, + 0.004167696926742792, + 0.006908314302563667, + 0.002640281105414033, + 0.004938727710396051, + 0.005410764832049608, + 0.004621243104338646, + 0.006043491419404745, + 0.00447905994951725, + 0.0069423639215528965, + 0.005553399678319693, + 0.00412736414000392, + 0.006106710992753506, + 0.005924854893237352, + 0.004723777994513512, + 0.009776853024959564, + 0.010967051610350609, + 0.011382711119949818, + 0.005230719689279795, + 0.010580433532595634, + 0.0124277388677001, + 0.011879093945026398, + 0.006686302367597818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.029819844290614128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635422024875879 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635422024875879, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635422024875879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11514846980571747, + 0.31632140278816223, + 0.05466778948903084, + 0.04352007061243057, + 0.07840093225240707, + 0.02424263209104538, + 0.05340128764510155, + 0.04587385430932045, + 0.07730469852685928, + 0.19111888110637665 + ], + "credit_map": [ + 0.010828472673892975, + 0.01183920819312334, + 0.018159672617912292, + 0.016061846166849136, + 0.014150967821478844, + 0.019336048513650894, + 0.009261500090360641, + 0.01551074255257845, + 0.022330448031425476, + 0.008928520604968071, + 0.1190343052148819, + 0.0120533537119627, + 0.007213254924863577, + 0.1317664533853531, + 0.007294255308806896, + 0.007700822781771421, + 0.007267712615430355, + 0.00591873237863183, + 0.006712844595313072, + 0.00770557252690196, + 0.003747710259631276, + 0.005607953295111656, + 0.009703783318400383, + 0.008003480732440948, + 0.005117027088999748, + 0.003810148686170578, + 0.005853865295648575, + 0.004610651172697544, + 0.005617588758468628, + 0.005594460293650627, + 0.0058668190613389015, + 0.007049510721117258, + 0.006568228360265493, + 0.00726939644664526, + 0.012116368860006332, + 0.0070749553851783276, + 0.00924938078969717, + 0.010033370926976204, + 0.01520987693220377, + 0.010879348032176495, + 0.002090661320835352, + 0.002002329332754016, + 0.0038439303170889616, + 0.002441516611725092, + 0.0018496413249522448, + 0.003602589713409543, + 0.005676936823874712, + 0.002735026879236102, + 0.0055099292658269405, + 0.003344009630382061, + 0.005648604594171047, + 0.0046180980280041695, + 0.004206529818475246, + 0.010411310009658337, + 0.006250506732612848, + 0.013412300497293472, + 0.0029390358831733465, + 0.002819584682583809, + 0.004479972179979086, + 0.010438886471092701, + 0.003421179251745343, + 0.0050314003601670265, + 0.010004359297454357, + 0.006739436648786068, + 0.008563278242945671, + 0.009149705059826374, + 0.012024789117276669, + 0.013153409585356712, + 0.006282174959778786, + 0.010228678584098816, + 0.01160637941211462, + 0.006296281702816486, + 0.021645834669470787, + 0.02638825960457325, + 0.020053686574101448, + 0.01824476569890976, + 0.03308230638504028, + 0.02702382020652294, + 0.029023759067058563, + 0.01565643772482872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.004635422024875879 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019392033573240042 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019392033573240042, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019392033573240042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06365078687667847, + 0.5785203576087952, + 0.02274210937321186, + 0.03988391160964966, + 0.021732598543167114, + 0.025530457496643066, + 0.03435457870364189, + 0.0716024860739708, + 0.04929589107632637, + 0.09268676489591599 + ], + "credit_map": [ + 0.0069231609813869, + 0.006745779886841774, + 0.009866737760603428, + 0.007031319197267294, + 0.006849933415651321, + 0.01143698114901781, + 0.004500582814216614, + 0.010296299122273922, + 0.0403306819498539, + 0.012755789794027805, + 0.22189657390117645, + 0.02061665989458561, + 0.01271969173103571, + 0.24603217840194702, + 0.01116102933883667, + 0.013007781468331814, + 0.0024752479512244463, + 0.0025750864297151566, + 0.003761891508474946, + 0.002373750787228346, + 0.00284490454941988, + 0.0037751730997115374, + 0.0019181693205609918, + 0.003017887705937028, + 0.004831031430512667, + 0.0037958945613354445, + 0.0053176432847976685, + 0.005428660195320845, + 0.004767287522554398, + 0.005533990915864706, + 0.004888429306447506, + 0.005320979282259941, + 0.002191022038459778, + 0.0016359258443117142, + 0.003218602854758501, + 0.0018257967894896865, + 0.002102190861478448, + 0.0027761925011873245, + 0.0036721036303788424, + 0.004310767166316509, + 0.001981654204428196, + 0.002163244178518653, + 0.0035010939463973045, + 0.0034624419640749693, + 0.0035476298071444035, + 0.002883430104702711, + 0.005152475088834763, + 0.002838487969711423, + 0.0030050750356167555, + 0.0029095320496708155, + 0.0032002581283450127, + 0.0033737088087946177, + 0.003172691911458969, + 0.004470333456993103, + 0.004146760329604149, + 0.010076221078634262, + 0.005338112823665142, + 0.004859233275055885, + 0.007275240961462259, + 0.01695025898516178, + 0.0063622817397117615, + 0.006228427402675152, + 0.0145860705524683, + 0.010002863593399525, + 0.004673860967159271, + 0.0035203532315790653, + 0.0069600618444383144, + 0.005921907722949982, + 0.006891020108014345, + 0.008306886069476604, + 0.006177438423037529, + 0.006844365503638983, + 0.008465046063065529, + 0.012199552729725838, + 0.014535263180732727, + 0.007580898702144623, + 0.013311583548784256, + 0.014204289764165878, + 0.013255823403596878, + 0.00913431216031313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.0019392033573240042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632378600537777 + ], + "opposite_logprob_tokens": [ + -3.6513237953186035 + ], + "logprob_reward": -0.02632378600537777, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632378600537777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06993863731622696, + 0.6016917824745178, + 0.018973801285028458, + 0.036797620356082916, + 0.04533292353153229, + 0.03143085911870003, + 0.02098383568227291, + 0.02931584231555462, + 0.07888377457857132, + 0.0666508898139 + ], + "credit_map": [ + 0.008662655018270016, + 0.006538013927638531, + 0.010910145938396454, + 0.008171997033059597, + 0.00694241002202034, + 0.008971933275461197, + 0.008120222948491573, + 0.011621261946856976, + 0.04190124198794365, + 0.012455849908292294, + 0.23192043602466583, + 0.020987220108509064, + 0.012237010523676872, + 0.25727352499961853, + 0.011631510220468044, + 0.013284988701343536, + 0.0019984019454568624, + 0.0016096469480544329, + 0.0025177595671266317, + 0.0026019979268312454, + 0.0012266404228284955, + 0.0021109895315021276, + 0.004034697078168392, + 0.0028736679814755917, + 0.004597840830683708, + 0.00400916114449501, + 0.004984902683645487, + 0.004613598342984915, + 0.005212468095123768, + 0.00414394261315465, + 0.004104834049940109, + 0.005130872130393982, + 0.005153345875442028, + 0.0032603403087705374, + 0.00712919794023037, + 0.0034751107450574636, + 0.0028779180720448494, + 0.006857447791844606, + 0.0035226624459028244, + 0.013056903146207333, + 0.002859602216631174, + 0.003810082795098424, + 0.004706392530351877, + 0.0029932239558547735, + 0.004187077283859253, + 0.005236234050244093, + 0.005070625338703394, + 0.002567622810602188, + 0.0017091869376599789, + 0.002186600584536791, + 0.0018157886806875467, + 0.0035665433388203382, + 0.002641557715833187, + 0.002790839411318302, + 0.002693163463845849, + 0.0035801557824015617, + 0.0018584660720080137, + 0.0020674329716712236, + 0.004334319848567247, + 0.005751037038862705, + 0.0019748860504478216, + 0.003635113826021552, + 0.005383167415857315, + 0.004311421420425177, + 0.0073057436384260654, + 0.006125340703874826, + 0.010679562576115131, + 0.009711365215480328, + 0.0059436289593577385, + 0.013536619953811169, + 0.016984768211841583, + 0.008596745319664478, + 0.007113907486200333, + 0.008542629890143871, + 0.006563584320247173, + 0.007522516883909702, + 0.010870713740587234, + 0.009872058406472206, + 0.01051945611834526, + 0.005646026227623224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.02632378600537777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted, extreme close-up captures a solid bat emblem, forged from cold, subtly scratched black metal, its outline imbued with tension. Against the backdrop of a blurred, rain-soaked, Gothic cityscape, flickering crimson and electric blue neon lights cast a shadow on the emblem's surface. The emblem begins to dissolve precariously from its sharp edges, its smooth metallic texture transforming into a thick, oily, inky liquid that slowly drips downward, forming plump droplets. Rendered with the exquisite clarity of ray tracing, each falling droplet resembles a miniature convex mirror, accurately reflecting and distorting the surrounding shimmering neon signs and cold city lights. The slow, eerie interplay of liquid flow and reflection creates a surreal visual spectacle set in a dark, ominous, and technologically futuristic atmosphere.", + "seed": 4522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_0/index_1422_seed_4522_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06929965927265584, + "vlm_logprob": -0.06929965927265584, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06929965927265584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22851724922657013 + ], + "opposite_logprob_tokens": [ + -1.6035172939300537 + ], + "logprob_reward": -0.22851724922657013, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22851724922657013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06632906943559647, + 0.06000138074159622, + 0.06516203284263611, + 0.17884166538715363, + 0.15869706869125366, + 0.3159759044647217, + 0.06626205891370773, + 0.03557957336306572, + 0.03114352747797966, + 0.022007709369063377 + ], + "credit_map": [ + 0.009553887881338596, + 0.006870589684695005, + 0.011670805513858795, + 0.008894686587154865, + 0.006032182835042477, + 0.006992454174906015, + 0.006629586219787598, + 0.00968487374484539, + 0.007699957583099604, + 0.009114079177379608, + 0.008271374739706516, + 0.008440693840384483, + 0.006585713941603899, + 0.005655746906995773, + 0.006926737260073423, + 0.007307081017643213, + 0.01001523807644844, + 0.008370532654225826, + 0.010800016112625599, + 0.013955296948552132, + 0.006238061469048262, + 0.003547748550772667, + 0.007512763142585754, + 0.004722374025732279, + 0.015346042811870575, + 0.025580748915672302, + 0.024354903027415276, + 0.023575082421302795, + 0.02637661062180996, + 0.01998724974691868, + 0.01769709773361683, + 0.02592393383383751, + 0.020698368549346924, + 0.01272321492433548, + 0.04624958708882332, + 0.02091626636683941, + 0.01043148897588253, + 0.014212263748049736, + 0.010619424283504486, + 0.022846456617116928, + 0.049325909465551376, + 0.017345426604151726, + 0.1024291068315506, + 0.01908753626048565, + 0.016961440443992615, + 0.015658551827073097, + 0.011049157939851284, + 0.0841187909245491, + 0.007871645502746105, + 0.007817582227289677, + 0.006446979008615017, + 0.008259487338364124, + 0.009616021066904068, + 0.0070456271059811115, + 0.00834638997912407, + 0.010858329012989998, + 0.004581928718835115, + 0.0030505647882819176, + 0.007326209452003241, + 0.0036842436529695988, + 0.0027137515135109425, + 0.002873998135328293, + 0.004446915350854397, + 0.006901959888637066, + 0.003907869569957256, + 0.002537543186917901, + 0.0076817236840724945, + 0.0035790428519248962, + 0.0028428894001990557, + 0.001907214056700468, + 0.002449291292577982, + 0.0062379539012908936, + 0.0022750007919967175, + 0.002783557865768671, + 0.0033529656939208508, + 0.002116525312885642, + 0.0019606573041528463, + 0.0020190582145005465, + 0.003870786167681217, + 0.0036291570868343115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.22851724922657013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308372363448143 + ], + "opposite_logprob_tokens": [ + -3.168083667755127 + ], + "logprob_reward": -0.04308372363448143, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308372363448143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.174149289727211, + 0.06927014887332916, + 0.13336583971977234, + 0.0718003511428833, + 0.08373980224132538, + 0.24006949365139008, + 0.0861465185880661, + 0.06823929399251938, + 0.031876448541879654, + 0.041342731565237045 + ], + "credit_map": [ + 0.02341502718627453, + 0.022662580013275146, + 0.02487659454345703, + 0.02556854672729969, + 0.02765038050711155, + 0.01173224113881588, + 0.015113997273147106, + 0.02312992326915264, + 0.008418116718530655, + 0.009994064457714558, + 0.009674117900431156, + 0.009941245429217815, + 0.006947727873921394, + 0.006556046195328236, + 0.01024654321372509, + 0.007492287550121546, + 0.020915240049362183, + 0.01254467386752367, + 0.023788556456565857, + 0.03361157327890396, + 0.010084148496389389, + 0.005675513297319412, + 0.015548228286206722, + 0.011197911575436592, + 0.008292662911117077, + 0.011035813950002193, + 0.012688994407653809, + 0.00863563735038042, + 0.012894258834421635, + 0.0051915268413722515, + 0.006574100814759731, + 0.006487356033176184, + 0.010539287701249123, + 0.015374820679426193, + 0.012165196239948273, + 0.008348748087882996, + 0.011433711275458336, + 0.006762695498764515, + 0.007540902588516474, + 0.011574435979127884, + 0.038998331874608994, + 0.027174489572644234, + 0.048348043113946915, + 0.024517418816685677, + 0.029692770913243294, + 0.01593165472149849, + 0.012762540951371193, + 0.04264424741268158, + 0.008381997235119343, + 0.008326238952577114, + 0.013021157123148441, + 0.011699443683028221, + 0.008576632477343082, + 0.009157703258097172, + 0.012555057182908058, + 0.01442828681319952, + 0.01154167391359806, + 0.005948656238615513, + 0.012417160905897617, + 0.008436826057732105, + 0.005475377663969994, + 0.004839036148041487, + 0.0071410806849598885, + 0.012439481914043427, + 0.005027832463383675, + 0.004295203369110823, + 0.004285032395273447, + 0.003752996912226081, + 0.0034186667762696743, + 0.0022491677664220333, + 0.0032737674191594124, + 0.0055737802758812904, + 0.006595435086637735, + 0.004604640416800976, + 0.005550404544919729, + 0.004746888764202595, + 0.005071818828582764, + 0.002957441844046116, + 0.004633702337741852, + 0.007182401139289141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.04308372363448143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1326688826084137, + 0.049058761447668076, + 0.02443227916955948, + 0.045363184064626694, + 0.08797972649335861, + 0.4356955289840698, + 0.10357681661844254, + 0.028512366116046906, + 0.0576939694583416, + 0.035018451511859894 + ], + "credit_map": [ + 0.017613839358091354, + 0.013186996802687645, + 0.019834095612168312, + 0.019324922934174538, + 0.016932815313339233, + 0.011811265721917152, + 0.011325674131512642, + 0.02263927273452282, + 0.007104265969246626, + 0.008663629181683064, + 0.006911768112331629, + 0.007342176511883736, + 0.007040318567305803, + 0.0029797269962728024, + 0.004203461576253176, + 0.004813418723642826, + 0.003011254360899329, + 0.0030774211045354605, + 0.0037046391516923904, + 0.0043821693398058414, + 0.0025327035691589117, + 0.0023165771272033453, + 0.0028635889757424593, + 0.0025439246091991663, + 0.00448283925652504, + 0.007368684746325016, + 0.006791435647755861, + 0.006870991084724665, + 0.006232072599232197, + 0.004308342467993498, + 0.0045613981783390045, + 0.004747422877699137, + 0.009369569830596447, + 0.016440775245428085, + 0.013141680508852005, + 0.011529900133609772, + 0.0171655360609293, + 0.005393537227064371, + 0.006135566625744104, + 0.008803159929811954, + 0.05526396632194519, + 0.03154556825757027, + 0.11347708106040955, + 0.03848521411418915, + 0.028598356992006302, + 0.021643565967679024, + 0.026889855042099953, + 0.11979193240404129, + 0.014302804134786129, + 0.01396004669368267, + 0.010814710520207882, + 0.008950646035373211, + 0.01574948988854885, + 0.010077097453176975, + 0.012858938425779343, + 0.01686309278011322, + 0.0037656507920473814, + 0.0024546256754547358, + 0.004990169312804937, + 0.004486554767936468, + 0.0016930963611230254, + 0.0021007091272622347, + 0.004639624152332544, + 0.004381936509162188, + 0.006814925000071526, + 0.0027074450626969337, + 0.01648837700486183, + 0.003954948857426643, + 0.003215709701180458, + 0.0042595942504704, + 0.0043718572705984116, + 0.015881117433309555, + 0.00546622509136796, + 0.003951943479478359, + 0.004135604482144117, + 0.004346227739006281, + 0.0040432452224195, + 0.0025636060163378716, + 0.004970507230609655, + 0.005541090387851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017130000051110983 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017130000051110983, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017130000051110983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27325543761253357, + 0.0368085578083992, + 0.04587702453136444, + 0.042005326598882675, + 0.05605974420905113, + 0.3391357958316803, + 0.0904608815908432, + 0.040714722126722336, + 0.04279492795467377, + 0.03288755565881729 + ], + "credit_map": [ + 0.04040567949414253, + 0.02707354724407196, + 0.03954034298658371, + 0.03512534126639366, + 0.035276155918836594, + 0.024108581244945526, + 0.023629622533917427, + 0.0480961874127388, + 0.0054442547261714935, + 0.00646188436076045, + 0.005521136801689863, + 0.004644996486604214, + 0.004797926172614098, + 0.0029486040584743023, + 0.0035204999148845673, + 0.0034692573826760054, + 0.0072762928903102875, + 0.004136047326028347, + 0.008484046906232834, + 0.011035432107746601, + 0.0029009452555328608, + 0.0027772258035838604, + 0.00521731935441494, + 0.004049714654684067, + 0.0039440663531422615, + 0.006163326092064381, + 0.006498402450233698, + 0.005589507520198822, + 0.006723993457853794, + 0.004180965479463339, + 0.004132381174713373, + 0.004772682674229145, + 0.00655400101095438, + 0.009702783077955246, + 0.009936003014445305, + 0.0064330557361245155, + 0.010589126497507095, + 0.0035143629647791386, + 0.0035929305013269186, + 0.0057374779134988785, + 0.0492108091711998, + 0.015571262687444687, + 0.10569512099027634, + 0.021575355902314186, + 0.012829530984163284, + 0.01705760881304741, + 0.013894092291593552, + 0.1033020094037056, + 0.012806604616343975, + 0.01322703156620264, + 0.00830048881471157, + 0.007816871628165245, + 0.015494479797780514, + 0.008028987795114517, + 0.010186811909079552, + 0.014599607326090336, + 0.005578169599175453, + 0.003798295743763447, + 0.006036210805177689, + 0.003588210791349411, + 0.004276171792298555, + 0.0034633104223757982, + 0.006424198392778635, + 0.007550158072263002, + 0.005181898362934589, + 0.0020166183821856976, + 0.011426448822021484, + 0.0026555676013231277, + 0.0022548853885382414, + 0.0035760030150413513, + 0.0035209469497203827, + 0.012162559665739536, + 0.0047941626980900764, + 0.0033768636640161276, + 0.003992985002696514, + 0.0046164775267243385, + 0.003626856952905655, + 0.0028401094023138285, + 0.004802040755748749, + 0.0048380629159510136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0017130000051110983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a solid bat emblem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the emblem made of black metal with a scratched texture and a tilted orientation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emblem transform from solid metal into a thick, oily, inky liquid that drips downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the falling droplets reflect and distort the surrounding crimson and electric blue neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, rain-soaked Gothic cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999258697032928 + ], + "opposite_logprob_tokens": [ + -2.6949925422668457 + ], + "logprob_reward": -0.06999258697032928, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999258697032928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13174785673618317, + 0.09290952235460281, + 0.08154618740081787, + 0.016521716490387917, + 0.0484720878303051, + 0.4586227834224701, + 0.07259386032819748, + 0.025310510769486427, + 0.050208043307065964, + 0.022067440673708916 + ], + "credit_map": [ + 0.020807674154639244, + 0.010987022891640663, + 0.019706066697835922, + 0.015662461519241333, + 0.016679957509040833, + 0.009802401065826416, + 0.012613285332918167, + 0.02548898197710514, + 0.01305192057043314, + 0.016070852056145668, + 0.010378318838775158, + 0.01564585231244564, + 0.013419577851891518, + 0.006286859046667814, + 0.010476873256266117, + 0.007579271215945482, + 0.013304962776601315, + 0.008065438829362392, + 0.015790970996022224, + 0.02071983925998211, + 0.005926829297095537, + 0.003265213221311569, + 0.009175256825983524, + 0.005297667812556028, + 0.001463854219764471, + 0.002432934707030654, + 0.0024207199458032846, + 0.0029281380120664835, + 0.00225178268738091, + 0.0015993262641131878, + 0.0015649229753762484, + 0.0018600381445139647, + 0.0050352816469967365, + 0.006874693091958761, + 0.01104048639535904, + 0.009088212624192238, + 0.004155911039561033, + 0.0027499531861394644, + 0.0036969599314033985, + 0.005830593407154083, + 0.06659812480211258, + 0.02708037756383419, + 0.15188579261302948, + 0.025816377252340317, + 0.020196184515953064, + 0.021860478445887566, + 0.016490526497364044, + 0.12869492173194885, + 0.00696134939789772, + 0.007029405329376459, + 0.008370102383196354, + 0.00785126630216837, + 0.0074762278236448765, + 0.008783604949712753, + 0.014378514140844345, + 0.0117433937266469, + 0.0031523292418569326, + 0.0018376627704128623, + 0.0033590684179216623, + 0.002683919621631503, + 0.0024712826125323772, + 0.0021176568698138, + 0.004986186511814594, + 0.004702404141426086, + 0.005796987563371658, + 0.002377298893406987, + 0.01628660224378109, + 0.004214268643409014, + 0.0026582665741443634, + 0.002878255909308791, + 0.0030567876528948545, + 0.012939570471644402, + 0.0023905981797724962, + 0.0014012011233717203, + 0.004234413616359234, + 0.0026295154821127653, + 0.0017011730233207345, + 0.002246924676001072, + 0.0035017880145460367, + 0.003961827140301466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.06999258697032928 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_tensor([4204, 4204])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_tensor([4204, 4204])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..400903bfbe124732b562169181492717e6e89158 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_tensor([4204, 4204])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, full-length shot, held at eye level, slowly pans to the left in an elegant circular trajectory, revealing a dancer at the center of a tranquil courtyard. The dancer, a woman with long, raven hair arranged in intricate classical curls, secured with a luminous jade hairpin, performs a mesmerizing choreography. She wears a flowing, gauzy silk Hanfu (Chinese traditional costume) in a pale, ethereal celadon green, its wide sleeves trailing behind her with every movement. Her expression is calm, her eyes half-closed, a look of graceful contemplation. The courtyard is framed by magnificent traditional Chinese architecture; dark-painted wooden pillars support a vast, gray-tiled roof with upturned eaves, while intricate latticed screens cast complex, ever-shifting shadows. The warm golden light of late afternoon bathes the landscape, capturing the delicate luster of the silk and highlighting the rich texture of the aged wood. Her movements are deliberately graceful, as she fluidly extends one arm, her wide sleeves unfurling like a painted scroll, followed by a slow, controlled rotation, her body a perfect counterweight as the silk billows around her, creating a breathtaking spectacle of movement and stillness. The atmosphere is one of profound serenity and timeless beauty, a living painting of elegance and architectural harmony.", + "seed": 4683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4683_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10940465060994028, + "vlm_logprob": -0.10940465060994028, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.10940465060994028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14153729379177094 + ], + "opposite_logprob_tokens": [ + -2.2665371894836426 + ], + "logprob_reward": -0.14153729379177094, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.14153729379177094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0761646181344986, + 0.04777256026864052, + 0.04119407385587692, + 0.07119809836149216, + 0.0381472110748291, + 0.5634489059448242, + 0.07163531333208084, + 0.03764956444501877, + 0.02091768942773342, + 0.03187188878655434 + ], + "credit_map": [ + 0.00794943980872631, + 0.014269109815359116, + 0.007444413844496012, + 0.00872967578470707, + 0.008552503772079945, + 0.008995916694402695, + 0.012313760817050934, + 0.007909809239208698, + 0.005512679927051067, + 0.007406229618936777, + 0.006257427390664816, + 0.0053445277735590935, + 0.007740824483335018, + 0.004958802368491888, + 0.004885722883045673, + 0.0056663500145077705, + 0.0045524886809289455, + 0.009121393784880638, + 0.005656805820763111, + 0.006487100385129452, + 0.0034970021806657314, + 0.0042269532568752766, + 0.004572227131575346, + 0.0030801056418567896, + 0.007008346263319254, + 0.014464357867836952, + 0.006224497687071562, + 0.01848364993929863, + 0.006638458464294672, + 0.006606778595596552, + 0.006560056935995817, + 0.005211954470723867, + 0.0034157163463532925, + 0.0057832542806863785, + 0.005042850971221924, + 0.005925206001847982, + 0.003026579739525914, + 0.0025299827102571726, + 0.007822670973837376, + 0.004600952845066786, + 0.05504557862877846, + 0.042394381016492844, + 0.07676546275615692, + 0.13786077499389648, + 0.03542148321866989, + 0.030478402972221375, + 0.12247446924448013, + 0.0630083829164505, + 0.00860598310828209, + 0.00759149482473731, + 0.008517537266016006, + 0.007205103989690542, + 0.008776627480983734, + 0.011151476763188839, + 0.011155642569065094, + 0.00863146223127842, + 0.0029156534001231194, + 0.0035819506738334894, + 0.003968667704612017, + 0.006432731635868549, + 0.0037559436168521643, + 0.00471147894859314, + 0.004676476586610079, + 0.007606666535139084, + 0.004339915234595537, + 0.0015567110385745764, + 0.0019337090197950602, + 0.001752822776325047, + 0.0023787592072039843, + 0.0020226677879691124, + 0.001525267492979765, + 0.005407838150858879, + 0.0034545843955129385, + 0.002572405617684126, + 0.004391115624457598, + 0.00573528790846467, + 0.005467602983117104, + 0.002003338187932968, + 0.0052517009899020195, + 0.0029958554077893496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.14153729379177094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486443758010864 + ], + "opposite_logprob_tokens": [ + -1.2236443758010864 + ], + "logprob_reward": -0.3486443758010864, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486443758010864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20431557297706604, + 0.1733517348766327, + 0.08989531546831131, + 0.05655164644122124, + 0.07322733849287033, + 0.08019512891769409, + 0.10811540484428406, + 0.07135435938835144, + 0.07947362959384918, + 0.0635199174284935 + ], + "credit_map": [ + 0.022917164489626884, + 0.04899187386035919, + 0.02579306811094284, + 0.019815988838672638, + 0.017861874774098396, + 0.02704515866935253, + 0.01994202844798565, + 0.021948402747511864, + 0.015475556254386902, + 0.021383333951234818, + 0.025321712717413902, + 0.01708105392754078, + 0.02147245593369007, + 0.022888507694005966, + 0.02936139516532421, + 0.020367711782455444, + 0.016316596418619156, + 0.009817629121243954, + 0.013134381733834743, + 0.012149343267083168, + 0.009262376464903355, + 0.010268167592585087, + 0.010262741707265377, + 0.008684077300131321, + 0.007628986146301031, + 0.011220546439290047, + 0.006459202617406845, + 0.012116322293877602, + 0.0040311384946107864, + 0.004720587283372879, + 0.005779668223112822, + 0.0045951916836202145, + 0.0067582870833575726, + 0.006727252621203661, + 0.004847733303904533, + 0.00906866230070591, + 0.007302159909158945, + 0.007226136513054371, + 0.021773993968963623, + 0.009523111395537853, + 0.007606057450175285, + 0.015830233693122864, + 0.014546767808496952, + 0.009994769468903542, + 0.00832937378436327, + 0.0075318654999136925, + 0.007398164365440607, + 0.008957888931035995, + 0.011354262940585613, + 0.01242142729461193, + 0.011712929233908653, + 0.011521423235535622, + 0.010031486861407757, + 0.016608471050858498, + 0.02138836868107319, + 0.013077023439109325, + 0.006757656112313271, + 0.008987919427454472, + 0.00916258804500103, + 0.011724845506250858, + 0.008956918492913246, + 0.0064690145663917065, + 0.009920930489897728, + 0.009374482557177544, + 0.007347621489316225, + 0.0075904191471636295, + 0.01181759126484394, + 0.019877443090081215, + 0.005736481864005327, + 0.009280913509428501, + 0.00950014591217041, + 0.008323009125888348, + 0.004971457179635763, + 0.004428823944181204, + 0.009883714839816093, + 0.010830812156200409, + 0.005631014704704285, + 0.009627570398151875, + 0.00921059213578701, + 0.008935932070016861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.3486443758010864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103692714124918 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.004103692714124918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103692714124918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14325666427612305, + 0.03993912413716316, + 0.05351179465651512, + 0.026697179302573204, + 0.021031329408288002, + 0.5126625299453735, + 0.11325637251138687, + 0.012175213545560837, + 0.03841286897659302, + 0.03905685991048813 + ], + "credit_map": [ + 0.015308075584471226, + 0.03289641812443733, + 0.008940706960856915, + 0.013392815366387367, + 0.01676063984632492, + 0.019446443766355515, + 0.01905263029038906, + 0.017458941787481308, + 0.003590891370549798, + 0.003694084705784917, + 0.004659675061702728, + 0.00427841953933239, + 0.003250658279284835, + 0.006575327832251787, + 0.006310087628662586, + 0.007579985074698925, + 0.010840397328138351, + 0.005479642655700445, + 0.01075560413300991, + 0.004514159634709358, + 0.003553170943632722, + 0.008580125868320465, + 0.004325380548834801, + 0.005463317036628723, + 0.003831506008282304, + 0.004692258313298225, + 0.0035609151236712933, + 0.0038973286282271147, + 0.0022348256316035986, + 0.0026109826285392046, + 0.0029277957510203123, + 0.002941568847745657, + 0.002199887065216899, + 0.0023112723138183355, + 0.0025157956406474113, + 0.004066163208335638, + 0.00223963032476604, + 0.0018569340463727713, + 0.003399674315005541, + 0.0024419750552624464, + 0.051655106246471405, + 0.03997837379574776, + 0.06916116923093796, + 0.12415186315774918, + 0.03237680718302727, + 0.028183676302433014, + 0.10937730967998505, + 0.057778287678956985, + 0.012360966764390469, + 0.01064294669777155, + 0.013583008199930191, + 0.01607278175652027, + 0.010864552110433578, + 0.018710309639573097, + 0.01885862462222576, + 0.012163199484348297, + 0.0018117493018507957, + 0.000983947655186057, + 0.0014815909089520574, + 0.0009777534287422895, + 0.001302518299780786, + 0.001974972430616617, + 0.0012923992471769452, + 0.002350283320993185, + 0.008544777520000935, + 0.0030812451150268316, + 0.0027794514317065477, + 0.003407679498195648, + 0.004081646911799908, + 0.0034744173754006624, + 0.003064386546611786, + 0.009979266673326492, + 0.0042436798103153706, + 0.004469704814255238, + 0.005087684839963913, + 0.006764897610992193, + 0.006839455105364323, + 0.0025802284944802523, + 0.005252727773040533, + 0.003818482393398881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.004103692714124918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971717953681946, + 0.11845598369836807, + 0.11758420616388321, + 0.057480134069919586, + 0.02784479409456253, + 0.36125513911247253, + 0.02807735465466976, + 0.032713647931814194, + 0.08198173344135284, + 0.08488981425762177 + ], + "credit_map": [ + 0.011349732056260109, + 0.00970317330211401, + 0.008083135820925236, + 0.009822588413953781, + 0.010654505342245102, + 0.011177638545632362, + 0.015381894074380398, + 0.013544507324695587, + 0.014197506941854954, + 0.010125143453478813, + 0.022058716043829918, + 0.00899787712842226, + 0.009936964139342308, + 0.018565721809864044, + 0.011844233609735966, + 0.022729812189936638, + 0.01805427111685276, + 0.014717258512973785, + 0.01564948260784149, + 0.011096050031483173, + 0.015820691362023354, + 0.016079511493444443, + 0.013326253741979599, + 0.012840690091252327, + 0.005487983580678701, + 0.008312487974762917, + 0.006480492651462555, + 0.010544768534600735, + 0.006079505663365126, + 0.008328708820044994, + 0.005502162501215935, + 0.006744024343788624, + 0.0029854446183890104, + 0.004017064813524485, + 0.002725091530010104, + 0.003916594199836254, + 0.003096819156780839, + 0.004027728922665119, + 0.003515198128297925, + 0.0035608510952442884, + 0.036053117364645004, + 0.027664804831147194, + 0.04901668801903725, + 0.08836938440799713, + 0.023003213107585907, + 0.020041700452566147, + 0.07666075229644775, + 0.040445491671562195, + 0.003312425222247839, + 0.0023067337460815907, + 0.002703579841181636, + 0.003677690401673317, + 0.0033819356467574835, + 0.0055595082230865955, + 0.004137578420341015, + 0.002997901290655136, + 0.002289676805958152, + 0.002516709268093109, + 0.004398301709443331, + 0.002766783582046628, + 0.0039480263367295265, + 0.003920966759324074, + 0.004080656915903091, + 0.00879252701997757, + 0.018628470599651337, + 0.006695501506328583, + 0.006313561927527189, + 0.006880669854581356, + 0.008626079186797142, + 0.0072449068538844585, + 0.005496667232364416, + 0.02209586650133133, + 0.008685043081641197, + 0.008190489374101162, + 0.012182201258838177, + 0.012986531481146812, + 0.016533536836504936, + 0.005897988565266132, + 0.011535176075994968, + 0.008878848515450954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864523932337761 + ], + "opposite_logprob_tokens": [ + -3.048645257949829 + ], + "logprob_reward": -0.04864523932337761, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864523932337761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3335055112838745, + 0.05216490849852562, + 0.056927166879177094, + 0.024267032742500305, + 0.05780189484357834, + 0.3114495277404785, + 0.032624248415231705, + 0.016514044255018234, + 0.021954506635665894, + 0.09279116988182068 + ], + "credit_map": [ + 0.03455231338739395, + 0.08788031339645386, + 0.01812518574297428, + 0.03069491684436798, + 0.033082470297813416, + 0.05282948911190033, + 0.0322648324072361, + 0.04407601058483124, + 0.005546822678297758, + 0.010195530951023102, + 0.007883207872509956, + 0.006314097438007593, + 0.005919909104704857, + 0.004789216443896294, + 0.0051891617476940155, + 0.00632696645334363, + 0.012595513835549355, + 0.006536056753247976, + 0.007073183543980122, + 0.006029622163623571, + 0.007260253187268972, + 0.008333319798111916, + 0.005317279603332281, + 0.0037819412536919117, + 0.0025164512917399406, + 0.003631139639765024, + 0.0029190797358751297, + 0.0038508358411490917, + 0.0029914823826402426, + 0.003409498604014516, + 0.002660940634086728, + 0.002287604147568345, + 0.005920927040278912, + 0.0048083094879984856, + 0.005133844446390867, + 0.005693287588655949, + 0.007751356810331345, + 0.00995474774390459, + 0.007867630571126938, + 0.010671799071133137, + 0.029566580429673195, + 0.023671643808484077, + 0.042766984552145004, + 0.07709034532308578, + 0.01937301829457283, + 0.017509695142507553, + 0.06654320657253265, + 0.034928057342767715, + 0.0036057569086551666, + 0.0046502165496349335, + 0.0032871440052986145, + 0.004253137856721878, + 0.004446651786565781, + 0.005733361467719078, + 0.0029131933115422726, + 0.0037347900215536356, + 0.0015700977528467774, + 0.001354158273898065, + 0.0018623810028657317, + 0.003389343386515975, + 0.0018017964903265238, + 0.0017237362917512655, + 0.001527396496385336, + 0.003285135142505169, + 0.002053195610642433, + 0.0028451301623135805, + 0.0028359696734696627, + 0.004273795522749424, + 0.0028349512722343206, + 0.002197546884417534, + 0.0025943838991224766, + 0.0023195340763777494, + 0.010515165515244007, + 0.009323536418378353, + 0.013725615106523037, + 0.016256660223007202, + 0.0166880302131176, + 0.00557563966140151, + 0.01273483969271183, + 0.007971678860485554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.04864523932337761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, full-length shot, held at eye level, slowly pans to the left in an elegant circular trajectory, revealing a dancer at the center of a tranquil courtyard. The dancer, a woman with long, raven hair arranged in intricate classical curls, secured with a luminous jade hairpin, performs a mesmerizing choreography. She wears a flowing, gauzy silk Hanfu (Chinese traditional costume) in a pale, ethereal celadon green, its wide sleeves trailing behind her with every movement. Her expression is calm, her eyes half-closed, a look of graceful contemplation. The courtyard is framed by magnificent traditional Chinese architecture; dark-painted wooden pillars support a vast, gray-tiled roof with upturned eaves, while intricate latticed screens cast complex, ever-shifting shadows. The warm golden light of late afternoon bathes the landscape, capturing the delicate luster of the silk and highlighting the rich texture of the aged wood. Her movements are deliberately graceful, as she fluidly extends one arm, her wide sleeves unfurling like a painted scroll, followed by a slow, controlled rotation, her body a perfect counterweight as the silk billows around her, creating a breathtaking spectacle of movement and stillness. The atmosphere is one of profound serenity and timeless beauty, a living painting of elegance and architectural harmony.", + "seed": 4684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4684_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11771801207214594, + "vlm_logprob": -0.11771801207214594, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.11771801207214594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2557409107685089 + ], + "opposite_logprob_tokens": [ + -1.6307408809661865 + ], + "logprob_reward": -0.2557409107685089, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2557409107685089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11575306206941605, + 0.04863415285944939, + 0.04956001788377762, + 0.05614537000656128, + 0.257053405046463, + 0.03795558586716652, + 0.1017179936170578, + 0.07927201688289642, + 0.1601007729768753, + 0.0938076600432396 + ], + "credit_map": [ + 0.013027239590883255, + 0.01027561817318201, + 0.01610591821372509, + 0.02313622646033764, + 0.01107429526746273, + 0.016203641891479492, + 0.014540526084601879, + 0.011389589868485928, + 0.005794048309326172, + 0.005399808753281832, + 0.006590643897652626, + 0.0064359200187027454, + 0.0057298531755805016, + 0.005086611025035381, + 0.007094590459018946, + 0.006502670701593161, + 0.004930553957819939, + 0.005782094318419695, + 0.008821600116789341, + 0.00994898285716772, + 0.004747949540615082, + 0.003433146746829152, + 0.0068426309153437614, + 0.005053053610026836, + 0.005361347924917936, + 0.006864432245492935, + 0.006031141150742769, + 0.007680374663323164, + 0.007148427423089743, + 0.005393314640969038, + 0.008586762472987175, + 0.009079568088054657, + 0.018268411979079247, + 0.011953513137996197, + 0.016046570613980293, + 0.027624743059277534, + 0.0191845390945673, + 0.07477571070194244, + 0.0575004443526268, + 0.031699445098638535, + 0.005220190156251192, + 0.00427388958632946, + 0.005746535491198301, + 0.0037377087865024805, + 0.004001809284090996, + 0.005675585940480232, + 0.005061503499746323, + 0.0042383610270917416, + 0.010297298431396484, + 0.006598957348614931, + 0.010880319401621819, + 0.010998478159308434, + 0.007597300224006176, + 0.009082324802875519, + 0.01783839613199234, + 0.028424911201000214, + 0.008178680203855038, + 0.011172330006957054, + 0.014034152030944824, + 0.0104721300303936, + 0.008567960001528263, + 0.009259863756597042, + 0.00926300697028637, + 0.008323888294398785, + 0.013651092536747456, + 0.008348832838237286, + 0.06129501760005951, + 0.032609276473522186, + 0.006788012571632862, + 0.007284567225724459, + 0.02211170643568039, + 0.008012257516384125, + 0.007322106976062059, + 0.006457450799643993, + 0.021363982930779457, + 0.014572112821042538, + 0.005326718091964722, + 0.02635270170867443, + 0.006964538712054491, + 0.005448047537356615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.2557409107685089 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813500463962555 + ], + "opposite_logprob_tokens": [ + -1.406350016593933 + ], + "logprob_reward": -0.2813500463962555, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813500463962555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07414030283689499, + 0.1411760449409485, + 0.05394691973924637, + 0.03760231286287308, + 0.0713646411895752, + 0.10579917579889297, + 0.11273600906133652, + 0.1284104436635971, + 0.1918783038854599, + 0.082945816218853 + ], + "credit_map": [ + 0.00656461575999856, + 0.007575539406388998, + 0.010535083711147308, + 0.012870320118963718, + 0.00950847752392292, + 0.01046445406973362, + 0.008803512901067734, + 0.007818295620381832, + 0.011470391415059566, + 0.00967208482325077, + 0.035801000893116, + 0.011841170489788055, + 0.010480974800884724, + 0.014412995427846909, + 0.007158737629652023, + 0.04033869132399559, + 0.004252294544130564, + 0.0066818795166909695, + 0.0060033174231648445, + 0.008926608599722385, + 0.009136941283941269, + 0.004239844623953104, + 0.009275257587432861, + 0.005430775694549084, + 0.003371425671502948, + 0.00283205509185791, + 0.00469798082485795, + 0.004822757560759783, + 0.00547547172755003, + 0.00390100060030818, + 0.00738133629783988, + 0.0051202853210270405, + 0.005061786621809006, + 0.004081688355654478, + 0.006888568866997957, + 0.009494038298726082, + 0.009557785466313362, + 0.01618010737001896, + 0.008662395179271698, + 0.011438264511525631, + 0.02195853926241398, + 0.01013136561959982, + 0.015651624649763107, + 0.009503469802439213, + 0.010137675330042839, + 0.012638255953788757, + 0.012411052361130714, + 0.01336719561368227, + 0.008466842584311962, + 0.014841808006167412, + 0.00986212957650423, + 0.019759899005293846, + 0.009825506247580051, + 0.017233343794941902, + 0.01753118261694908, + 0.015215301886200905, + 0.017928186804056168, + 0.01369456760585308, + 0.016721900552511215, + 0.017625264823436737, + 0.009124831296503544, + 0.01959853805601597, + 0.014192193746566772, + 0.019524965435266495, + 0.015924090519547462, + 0.021570144221186638, + 0.015775343403220177, + 0.02906724438071251, + 0.014052076265215874, + 0.011786919087171555, + 0.0640449970960617, + 0.019657500088214874, + 0.00807979516685009, + 0.008020621724426746, + 0.016545189544558525, + 0.01179015263915062, + 0.007067643105983734, + 0.020041344687342644, + 0.006825319025665522, + 0.004575752187520266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.2813500463962555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032007675617933273 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032007675617933273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032007675617933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11850016564130783, + 0.19293691217899323, + 0.0398075245320797, + 0.030816439539194107, + 0.13217267394065857, + 0.06435782462358475, + 0.04991910234093666, + 0.08238942921161652, + 0.1986381709575653, + 0.09046176075935364 + ], + "credit_map": [ + 0.015499156899750233, + 0.010750193148851395, + 0.014452354982495308, + 0.01787167228758335, + 0.015203380957245827, + 0.021374179050326347, + 0.01148501317948103, + 0.011864220723509789, + 0.012940710410475731, + 0.013085910119116306, + 0.048034559935331345, + 0.013876023702323437, + 0.020885072648525238, + 0.017572790384292603, + 0.015050121583044529, + 0.051491737365722656, + 0.003959613852202892, + 0.0042933514341712, + 0.004696935880929232, + 0.0036329561844468117, + 0.00504627451300621, + 0.006347777787595987, + 0.007622674107551575, + 0.004207948222756386, + 0.0028485306538641453, + 0.004298360086977482, + 0.0030996978748589754, + 0.00429373886436224, + 0.0034626591950654984, + 0.0036913342773914337, + 0.005980615969747305, + 0.003141505178064108, + 0.014937830157577991, + 0.011592867784202099, + 0.01337728276848793, + 0.024690721184015274, + 0.018767518922686577, + 0.015331595204770565, + 0.019729241728782654, + 0.013745620846748352, + 0.018908824771642685, + 0.009418517351150513, + 0.008252538740634918, + 0.005144279450178146, + 0.004787817597389221, + 0.006511119659990072, + 0.005220051389187574, + 0.00611467519775033, + 0.005595197901129723, + 0.004207816906273365, + 0.006191533990204334, + 0.004197622183710337, + 0.005452982150018215, + 0.004911921452730894, + 0.0066188061609864235, + 0.012743224389851093, + 0.007785082329064608, + 0.016933389008045197, + 0.013007566332817078, + 0.013502256944775581, + 0.008064349181950092, + 0.00648161256685853, + 0.007270363159477711, + 0.009344808757305145, + 0.009568250738084316, + 0.00700595835223794, + 0.02143784612417221, + 0.0606946237385273, + 0.010877709835767746, + 0.006762744393199682, + 0.06678317487239838, + 0.015507873147726059, + 0.007701290305703878, + 0.008793103508651257, + 0.017588002607226372, + 0.016836518421769142, + 0.007108889985829592, + 0.020795591175556183, + 0.006786695681512356, + 0.004851671401411295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.0032007675617933273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14040657877922058, + 0.11930938065052032, + 0.07532831281423569, + 0.0750458613038063, + 0.08943869173526764, + 0.03297858312726021, + 0.058663446456193924, + 0.05663185566663742, + 0.28367310762405396, + 0.06852424889802933 + ], + "credit_map": [ + 0.013220935128629208, + 0.016493719071149826, + 0.017195669934153557, + 0.021352749317884445, + 0.018170295283198357, + 0.026816679164767265, + 0.013400400057435036, + 0.01375612523406744, + 0.011443273164331913, + 0.008077430538833141, + 0.031890545040369034, + 0.008655356243252754, + 0.006158856209367514, + 0.010561038739979267, + 0.007776996586471796, + 0.03474588319659233, + 0.0072900294326245785, + 0.008906364440917969, + 0.012632764875888824, + 0.014535443857312202, + 0.007326136808842421, + 0.005368608050048351, + 0.010801847092807293, + 0.008467121049761772, + 0.008568988181650639, + 0.017346682026982307, + 0.008321637287735939, + 0.009895270690321922, + 0.00617092614993453, + 0.00840127281844616, + 0.006083557847887278, + 0.010257521644234657, + 0.005901644472032785, + 0.0053091193549335, + 0.006171439308673143, + 0.009416639804840088, + 0.009666047058999538, + 0.025461344048380852, + 0.01725759543478489, + 0.010254868306219578, + 0.005410303827375174, + 0.00332242320291698, + 0.004333570133894682, + 0.002676083706319332, + 0.005600121337920427, + 0.00477311247959733, + 0.003792401636019349, + 0.0030705658718943596, + 0.0066134086810052395, + 0.0029710361268371344, + 0.008475551381707191, + 0.0040247696451842785, + 0.008974708616733551, + 0.006352188531309366, + 0.008107795380055904, + 0.013143982738256454, + 0.0069514173083007336, + 0.006429805886000395, + 0.007543155457824469, + 0.00755828944966197, + 0.004334416706115007, + 0.010601542890071869, + 0.005099148023873568, + 0.008114083670079708, + 0.015016335994005203, + 0.008527177385985851, + 0.022463800385594368, + 0.07085463404655457, + 0.018170511350035667, + 0.010727095417678356, + 0.11259844154119492, + 0.025315087288618088, + 0.004977811127901077, + 0.008386694826185703, + 0.009262279607355595, + 0.010908108204603195, + 0.013321940787136555, + 0.0075928219594061375, + 0.007678968366235495, + 0.006395627278834581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005250236950814724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304809868335724 + ], + "opposite_logprob_tokens": [ + -3.168048143386841 + ], + "logprob_reward": -0.04304809868335724, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304809868335724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15603524446487427, + 0.09478495270013809, + 0.061801351606845856, + 0.09356362372636795, + 0.11790727823972702, + 0.09171916544437408, + 0.04487188905477524, + 0.07489050924777985, + 0.12459002435207367, + 0.1398359090089798 + ], + "credit_map": [ + 0.017513716593384743, + 0.02457636222243309, + 0.018878834322094917, + 0.024028055369853973, + 0.018035439774394035, + 0.018819160759449005, + 0.016429360955953598, + 0.017754314467310905, + 0.006652734708040953, + 0.007499320432543755, + 0.022729555144906044, + 0.008229237049818039, + 0.008195239119231701, + 0.008876772597432137, + 0.007594863418489695, + 0.02500723861157894, + 0.005888618528842926, + 0.0068032280541956425, + 0.008886427618563175, + 0.006373452488332987, + 0.006025544833391905, + 0.009380071423947811, + 0.00907302089035511, + 0.009370987303555012, + 0.011633757501840591, + 0.014600942842662334, + 0.011153084225952625, + 0.012179925106465816, + 0.012966206297278404, + 0.009093144908547401, + 0.01257250364869833, + 0.00936405174434185, + 0.013488298282027245, + 0.011477428488433361, + 0.012872375547885895, + 0.02166101709008217, + 0.02144673839211464, + 0.011681441217660904, + 0.012296795845031738, + 0.012983188033103943, + 0.024121597409248352, + 0.012639356777071953, + 0.012635580264031887, + 0.008506792597472668, + 0.008767274208366871, + 0.009135776199400425, + 0.008026291616261005, + 0.007886494509875774, + 0.006088460795581341, + 0.004253907594829798, + 0.0037557657342404127, + 0.00452857930213213, + 0.004975671414285898, + 0.004433429799973965, + 0.007409100420773029, + 0.00942697562277317, + 0.010262049734592438, + 0.00818508118391037, + 0.007671378552913666, + 0.011267083697021008, + 0.009304297156631947, + 0.009315576404333115, + 0.010377436876296997, + 0.008507605642080307, + 0.008356613107025623, + 0.013418673537671566, + 0.01919533684849739, + 0.019965440034866333, + 0.007447496987879276, + 0.004487366881221533, + 0.04080841317772865, + 0.010910684242844582, + 0.011322973296046257, + 0.010166415013372898, + 0.029094312340021133, + 0.015332364477217197, + 0.015249032527208328, + 0.03593670204281807, + 0.012274865061044693, + 0.01045924611389637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.04304809868335724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, full-length shot, held at eye level, slowly pans to the left in an elegant circular trajectory, revealing a dancer at the center of a tranquil courtyard. The dancer, a woman with long, raven hair arranged in intricate classical curls, secured with a luminous jade hairpin, performs a mesmerizing choreography. She wears a flowing, gauzy silk Hanfu (Chinese traditional costume) in a pale, ethereal celadon green, its wide sleeves trailing behind her with every movement. Her expression is calm, her eyes half-closed, a look of graceful contemplation. The courtyard is framed by magnificent traditional Chinese architecture; dark-painted wooden pillars support a vast, gray-tiled roof with upturned eaves, while intricate latticed screens cast complex, ever-shifting shadows. The warm golden light of late afternoon bathes the landscape, capturing the delicate luster of the silk and highlighting the rich texture of the aged wood. Her movements are deliberately graceful, as she fluidly extends one arm, her wide sleeves unfurling like a painted scroll, followed by a slow, controlled rotation, her body a perfect counterweight as the silk billows around her, creating a breathtaking spectacle of movement and stillness. The atmosphere is one of profound serenity and timeless beauty, a living painting of elegance and architectural harmony.", + "seed": 4685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4685_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19417582624591886, + "vlm_logprob": -0.19417582624591886, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.19417582624591886, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.141100212931633 + ], + "opposite_logprob_tokens": [ + -2.2661001682281494 + ], + "logprob_reward": -0.141100212931633, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.141100212931633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22271941602230072, + 0.1041141077876091, + 0.12829554080963135, + 0.05157949775457382, + 0.14812549948692322, + 0.04158703237771988, + 0.057431455701589584, + 0.12465697526931763, + 0.07876818627119064, + 0.042722247540950775 + ], + "credit_map": [ + 0.030452663078904152, + 0.047178491950035095, + 0.028000615537166595, + 0.03226471319794655, + 0.01834086887538433, + 0.02126900851726532, + 0.016672538593411446, + 0.028540529310703278, + 0.017677931115031242, + 0.00925356149673462, + 0.014922603033483028, + 0.013405591249465942, + 0.012462791055440903, + 0.0077251554466784, + 0.01364297978579998, + 0.015023489482700825, + 0.01349622942507267, + 0.00619455985724926, + 0.04089498519897461, + 0.010016439482569695, + 0.008064384572207928, + 0.007862197235226631, + 0.010466353967785835, + 0.03130039572715759, + 0.0074846623465418816, + 0.010129381902515888, + 0.004889974370598793, + 0.008143724873661995, + 0.005980847403407097, + 0.004397000186145306, + 0.005783285014331341, + 0.00477062352001667, + 0.015981415286660194, + 0.006902312859892845, + 0.05310899764299393, + 0.007294993381947279, + 0.012401290237903595, + 0.008902125060558319, + 0.006474917754530907, + 0.03705945238471031, + 0.007855541072785854, + 0.004970044828951359, + 0.006570995785295963, + 0.005161148030310869, + 0.004454270005226135, + 0.00570853753015399, + 0.003244454739615321, + 0.003622036427259445, + 0.006338945589959621, + 0.005641589406877756, + 0.005061986856162548, + 0.009976117871701717, + 0.007693277671933174, + 0.007911166176199913, + 0.009987900964915752, + 0.0048204706981778145, + 0.016963375732302666, + 0.024443931877613068, + 0.0101297153159976, + 0.012576042674481869, + 0.015268076211214066, + 0.01481662318110466, + 0.019482987001538277, + 0.010976218618452549, + 0.010719980113208294, + 0.010056988336145878, + 0.010304681025445461, + 0.007116041611880064, + 0.010656671598553658, + 0.009580986574292183, + 0.008763094432651997, + 0.011569743975996971, + 0.006807514000684023, + 0.007306680548936129, + 0.005094495136290789, + 0.003924797289073467, + 0.0047346861101686954, + 0.008082516491413116, + 0.0029762685298919678, + 0.003795287339016795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.141100212931633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578127384185791 + ], + "opposite_logprob_tokens": [ + -0.6328127384185791 + ], + "logprob_reward": -0.7578127384185791, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578127384185791, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22200748324394226, + 0.09275262802839279, + 0.06372515112161636, + 0.03640064597129822, + 0.10037142038345337, + 0.07335072755813599, + 0.10673150420188904, + 0.07785998284816742, + 0.16596078872680664, + 0.0608396902680397 + ], + "credit_map": [ + 0.023827945813536644, + 0.05914728343486786, + 0.02054041624069214, + 0.025885161012411118, + 0.02400248870253563, + 0.027231039479374886, + 0.011749287135899067, + 0.029623856768012047, + 0.0165040735155344, + 0.009575278498232365, + 0.011248716153204441, + 0.008129114285111427, + 0.008737345226109028, + 0.009186158888041973, + 0.009205699898302555, + 0.020166238769888878, + 0.009286840446293354, + 0.004376403521746397, + 0.008355990052223206, + 0.006395915523171425, + 0.008574268780648708, + 0.004707271698862314, + 0.006487770937383175, + 0.015540684573352337, + 0.003962073475122452, + 0.006382453721016645, + 0.005103904753923416, + 0.006224235985428095, + 0.003667999990284443, + 0.002959784120321274, + 0.004443753976374865, + 0.0036564406473189592, + 0.01012862753123045, + 0.006091042421758175, + 0.03080492652952671, + 0.006495634559541941, + 0.008685596287250519, + 0.006614441517740488, + 0.012506388127803802, + 0.01904476247727871, + 0.00929119624197483, + 0.013739990070462227, + 0.010054206475615501, + 0.013519012369215488, + 0.009198816493153572, + 0.0077925920486450195, + 0.004773167427629232, + 0.004981752019375563, + 0.015816399827599525, + 0.017189547419548035, + 0.010969963856041431, + 0.016089145094156265, + 0.014340322464704514, + 0.013734751380980015, + 0.010526943020522594, + 0.008064426481723785, + 0.012612382881343365, + 0.015340411104261875, + 0.007472141645848751, + 0.007815997116267681, + 0.009237103164196014, + 0.009640166535973549, + 0.007101865950971842, + 0.008639915846288204, + 0.02042762190103531, + 0.018336063250899315, + 0.01586894318461418, + 0.021784573793411255, + 0.031354762613773346, + 0.016408473253250122, + 0.025383006781339645, + 0.016397343948483467, + 0.01463934313505888, + 0.0045583113096654415, + 0.014835511334240437, + 0.004942456725984812, + 0.007384548429399729, + 0.00455006118863821, + 0.003855664050206542, + 0.006073791533708572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.7578127384185791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20646411180496216, + 0.07636355608701706, + 0.033317312598228455, + 0.03675095736980438, + 0.07360396534204483, + 0.2522467076778412, + 0.13257533311843872, + 0.04947856813669205, + 0.07311224192380905, + 0.06608721613883972 + ], + "credit_map": [ + 0.024840492755174637, + 0.06401381641626358, + 0.01770196296274662, + 0.023827258497476578, + 0.023403434082865715, + 0.02123677358031273, + 0.013668456114828587, + 0.017771925777196884, + 0.011940781027078629, + 0.009016039781272411, + 0.011630017310380936, + 0.008694387972354889, + 0.009067676030099392, + 0.006324045360088348, + 0.008466392755508423, + 0.01122421957552433, + 0.005112755112349987, + 0.0029390680138021708, + 0.0041718874126672745, + 0.003882773919031024, + 0.0043011740781366825, + 0.002968924818560481, + 0.00413666432723403, + 0.005804064683616161, + 0.0033463144209235907, + 0.004732619505375624, + 0.004588151816278696, + 0.0077125621028244495, + 0.0047555770725011826, + 0.002748750150203705, + 0.004009733442217112, + 0.004857251886278391, + 0.009736428037285805, + 0.0057701473124325275, + 0.021171974018216133, + 0.0046008569188416, + 0.0055963522754609585, + 0.004382826387882233, + 0.00464177830144763, + 0.017703598365187645, + 0.02463371679186821, + 0.01242372952401638, + 0.07878832519054413, + 0.05700152739882469, + 0.012812776491045952, + 0.0334174819290638, + 0.01988562010228634, + 0.013283506967127323, + 0.017184434458613396, + 0.01843613013625145, + 0.01285265851765871, + 0.017916863784193993, + 0.02122616581618786, + 0.0168811846524477, + 0.01491798460483551, + 0.01315990649163723, + 0.010586174204945564, + 0.0050987922586500645, + 0.003988842945545912, + 0.0040531461127102375, + 0.007967483252286911, + 0.005585397128015757, + 0.00672887871041894, + 0.005469851195812225, + 0.008307787589728832, + 0.013769041746854782, + 0.009218249469995499, + 0.008608411997556686, + 0.01002122089266777, + 0.008249865844845772, + 0.00797861535102129, + 0.00695904390886426, + 0.012126400135457516, + 0.005946947727352381, + 0.01906370371580124, + 0.006304933689534664, + 0.007426206022500992, + 0.005940769799053669, + 0.0042874799109995365, + 0.004990773741155863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.006741635967046022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031894787680357695 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.0031894787680357695, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031894787680357695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1738216131925583, + 0.09729961305856705, + 0.030653128400444984, + 0.07342206686735153, + 0.16700434684753418, + 0.062946617603302, + 0.18937791883945465, + 0.06082046031951904, + 0.09163400530815125, + 0.05302026495337486 + ], + "credit_map": [ + 0.018352780491113663, + 0.03943001478910446, + 0.01767171546816826, + 0.01926049217581749, + 0.019445786252617836, + 0.019100407138466835, + 0.014907421544194221, + 0.025652989745140076, + 0.008955078199505806, + 0.010132962837815285, + 0.009034978225827217, + 0.007723236922174692, + 0.009857998229563236, + 0.011969252489507198, + 0.023163244128227234, + 0.01646285690367222, + 0.0038363311905413866, + 0.0048234728164970875, + 0.0024500577710568905, + 0.004523316398262978, + 0.0035256734117865562, + 0.0039886790327727795, + 0.00422993628308177, + 0.003275657771155238, + 0.012965476140379906, + 0.00833126250654459, + 0.008527331054210663, + 0.009940432384610176, + 0.00850227102637291, + 0.006261157803237438, + 0.009670943953096867, + 0.009223183616995811, + 0.024586329236626625, + 0.009153914637863636, + 0.043740615248680115, + 0.009410991333425045, + 0.01407570019364357, + 0.013320972211658955, + 0.010851345025002956, + 0.041864462196826935, + 0.007498030085116625, + 0.006427271291613579, + 0.016548030078411102, + 0.011956230737268925, + 0.004409246612340212, + 0.0075465841218829155, + 0.004473383538424969, + 0.004087838344275951, + 0.0254722498357296, + 0.02780173532664776, + 0.01205799076706171, + 0.036740511655807495, + 0.02426976151764393, + 0.023551134392619133, + 0.027943627908825874, + 0.011540889739990234, + 0.009085732512176037, + 0.010961037129163742, + 0.007078456226736307, + 0.0052791074849665165, + 0.006190963089466095, + 0.006580275483429432, + 0.007711437996476889, + 0.00793344434350729, + 0.011120844632387161, + 0.016144081950187683, + 0.013139396905899048, + 0.010194704867899418, + 0.009238285012543201, + 0.012621786445379257, + 0.008239991031587124, + 0.010934919118881226, + 0.005687654484063387, + 0.007952096872031689, + 0.007959532551467419, + 0.006045158952474594, + 0.010388312861323357, + 0.006108542438596487, + 0.0037688049487769604, + 0.00511015672236681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0031894787680357695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203506514430046 + ], + "opposite_logprob_tokens": [ + -2.812035083770752 + ], + "logprob_reward": -0.06203506514430046, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203506514430046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5576273202896118, + 0.042076922953128815, + 0.04367988556623459, + 0.025906391441822052, + 0.09184100478887558, + 0.03616739809513092, + 0.07532322406768799, + 0.06265973299741745, + 0.025298316031694412, + 0.039419811218976974 + ], + "credit_map": [ + 0.05510881543159485, + 0.15462128818035126, + 0.06478950381278992, + 0.05579432472586632, + 0.061960797756910324, + 0.05326785892248154, + 0.020970584824681282, + 0.09111417829990387, + 0.006790350191295147, + 0.004257737658917904, + 0.006784407421946526, + 0.004048856440931559, + 0.004478940274566412, + 0.005493737757205963, + 0.004467116203159094, + 0.005755776073783636, + 0.006249013356864452, + 0.0029584981966763735, + 0.009541123174130917, + 0.0029739649035036564, + 0.004287865478545427, + 0.003520496189594269, + 0.0044120256789028645, + 0.009736895561218262, + 0.0035238845739513636, + 0.0030110832303762436, + 0.004399564117193222, + 0.0025196080096066, + 0.0036545286420732737, + 0.0026893378235399723, + 0.0023335653822869062, + 0.003774819429963827, + 0.012496508657932281, + 0.006548666395246983, + 0.026076506823301315, + 0.005495036020874977, + 0.00870911218225956, + 0.005776544101536274, + 0.004096492193639278, + 0.022642137482762337, + 0.008247853256762028, + 0.006700893398374319, + 0.004752421285957098, + 0.004252503626048565, + 0.0030873948708176613, + 0.0036494936794042587, + 0.0031872615218162537, + 0.0022895748261362314, + 0.017408236861228943, + 0.008323099464178085, + 0.013844233937561512, + 0.005908953957259655, + 0.007536182180047035, + 0.009696001186966896, + 0.006066849455237389, + 0.006539665628224611, + 0.007894469425082207, + 0.009400845505297184, + 0.006339312996715307, + 0.0062023065984249115, + 0.009372081607580185, + 0.00785539485514164, + 0.007140369154512882, + 0.008454956114292145, + 0.0034043407067656517, + 0.0025803784374147654, + 0.0028877679724246264, + 0.0028599826619029045, + 0.0043325345031917095, + 0.0026537347584962845, + 0.0028588653076440096, + 0.0037207109853625298, + 0.0039909593760967255, + 0.004235319793224335, + 0.005625592079013586, + 0.0060920268297195435, + 0.003838822478428483, + 0.007661120500415564, + 0.00367284775711596, + 0.004303121939301491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.06203506514430046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, full-length shot, held at eye level, slowly pans to the left in an elegant circular trajectory, revealing a dancer at the center of a tranquil courtyard. The dancer, a woman with long, raven hair arranged in intricate classical curls, secured with a luminous jade hairpin, performs a mesmerizing choreography. She wears a flowing, gauzy silk Hanfu (Chinese traditional costume) in a pale, ethereal celadon green, its wide sleeves trailing behind her with every movement. Her expression is calm, her eyes half-closed, a look of graceful contemplation. The courtyard is framed by magnificent traditional Chinese architecture; dark-painted wooden pillars support a vast, gray-tiled roof with upturned eaves, while intricate latticed screens cast complex, ever-shifting shadows. The warm golden light of late afternoon bathes the landscape, capturing the delicate luster of the silk and highlighting the rich texture of the aged wood. Her movements are deliberately graceful, as she fluidly extends one arm, her wide sleeves unfurling like a painted scroll, followed by a slow, controlled rotation, her body a perfect counterweight as the silk billows around her, creating a breathtaking spectacle of movement and stillness. The atmosphere is one of profound serenity and timeless beauty, a living painting of elegance and architectural harmony.", + "seed": 4686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_10/index_4204_seed_4686_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07703218348324299, + "vlm_logprob": -0.07703218348324299, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.07703218348324299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14649875462055206 + ], + "opposite_logprob_tokens": [ + -2.146498680114746 + ], + "logprob_reward": -0.14649875462055206, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14649875462055206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10547006875276566, + 0.05739862471818924, + 0.07703377306461334, + 0.16801482439041138, + 0.17256508767604828, + 0.10128670185804367, + 0.0613233707845211, + 0.10668148845434189, + 0.08432088792324066, + 0.06590517610311508 + ], + "credit_map": [ + 0.011043339967727661, + 0.008747683838009834, + 0.020215269178152084, + 0.010628842748701572, + 0.01509764976799488, + 0.010757926851511002, + 0.011519117280840874, + 0.01746024191379547, + 0.008141783997416496, + 0.0064129941165447235, + 0.010200953111052513, + 0.007681429851800203, + 0.008614526130259037, + 0.005557263735681772, + 0.005246574524790049, + 0.005543098319321871, + 0.007209934759885073, + 0.008610931225121021, + 0.019361069425940514, + 0.011919877491891384, + 0.01051301322877407, + 0.0038426024839282036, + 0.007692708633840084, + 0.007883638143539429, + 0.010926182381808758, + 0.01712919771671295, + 0.015364501625299454, + 0.06338080018758774, + 0.01278604194521904, + 0.0068191965110599995, + 0.02514314465224743, + 0.016465749591588974, + 0.020134424790740013, + 0.008493958041071892, + 0.046702198684215546, + 0.017417708411812782, + 0.010082713328301907, + 0.009465271607041359, + 0.027651354670524597, + 0.0326174721121788, + 0.008486030623316765, + 0.014257494360208511, + 0.014774552546441555, + 0.012102248147130013, + 0.016043810173869133, + 0.017306149005889893, + 0.011334419250488281, + 0.006981995422393084, + 0.008075927384197712, + 0.0051799616776406765, + 0.009317883290350437, + 0.006610519718378782, + 0.007835977710783482, + 0.00955995824187994, + 0.009546208195388317, + 0.0051969364285469055, + 0.013901564292609692, + 0.012115191668272018, + 0.01880035735666752, + 0.011064575053751469, + 0.01010824739933014, + 0.011988661251962185, + 0.01029161922633648, + 0.01841127872467041, + 0.009976213797926903, + 0.012221608310937881, + 0.00920839048922062, + 0.012937065213918686, + 0.012666360475122929, + 0.006875133141875267, + 0.014430675655603409, + 0.006005441769957542, + 0.006661350838840008, + 0.006076403893530369, + 0.017241666093468666, + 0.004922983702272177, + 0.006618245970457792, + 0.005675096996128559, + 0.004258421715348959, + 0.014451004564762115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.14649875462055206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428145170211792 + ], + "opposite_logprob_tokens": [ + -2.0178146362304688 + ], + "logprob_reward": -0.1428145170211792, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428145170211792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046089399605989456, + 0.0618295818567276, + 0.058355238288640976, + 0.0498703308403492, + 0.17684194445610046, + 0.09499883651733398, + 0.05312714725732803, + 0.09205145388841629, + 0.2434481829404831, + 0.12338785082101822 + ], + "credit_map": [ + 0.004038352984935045, + 0.0035594033543020487, + 0.005589839071035385, + 0.00545459846034646, + 0.011462966911494732, + 0.004262436181306839, + 0.006723729893565178, + 0.004998077172785997, + 0.00802598986774683, + 0.004807197023183107, + 0.014596477150917053, + 0.013339429162442684, + 0.010099966078996658, + 0.002767466474324465, + 0.005320423748344183, + 0.0028726423624902964, + 0.0057626813650131226, + 0.006331694778054953, + 0.00690847123041749, + 0.019567551091313362, + 0.005547504872083664, + 0.002412307308986783, + 0.0069164494052529335, + 0.004908583126962185, + 0.0036052034702152014, + 0.0049002706073224545, + 0.005184789653867483, + 0.017760630697011948, + 0.005935551133006811, + 0.0021096118725836277, + 0.006547214463353157, + 0.003827061038464308, + 0.019776728004217148, + 0.008934547193348408, + 0.05018923059105873, + 0.016163816675543785, + 0.010379524901509285, + 0.010061880573630333, + 0.026815740391612053, + 0.034520480781793594, + 0.01330910436809063, + 0.012777860276401043, + 0.010164174251258373, + 0.0077645001001656055, + 0.022571958601474762, + 0.011783124879002571, + 0.006824769079685211, + 0.009803342632949352, + 0.007205065805464983, + 0.004925540182739496, + 0.009485790506005287, + 0.006087061017751694, + 0.006023834925144911, + 0.009563172236084938, + 0.00515043456107378, + 0.004686254542320967, + 0.011031596921384335, + 0.010031353682279587, + 0.018676191568374634, + 0.008212383836507797, + 0.01035587303340435, + 0.012648742645978928, + 0.007103350479155779, + 0.013991966843605042, + 0.017705371603369713, + 0.02164481021463871, + 0.0642630085349083, + 0.028829077258706093, + 0.03915226459503174, + 0.017367469146847725, + 0.018531326204538345, + 0.03595488891005516, + 0.013854442164301872, + 0.009612164460122585, + 0.04061312600970268, + 0.005822788458317518, + 0.011643446050584316, + 0.01042497344315052, + 0.006442899815738201, + 0.024974022060632706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.1428145170211792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022025867365300655 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022025867365300655, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022025867365300655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09658379852771759, + 0.06126245856285095, + 0.09612987190485, + 0.1027471274137497, + 0.13433049619197845, + 0.08771726489067078, + 0.047734785825014114, + 0.1068255826830864, + 0.04568089544773102, + 0.2209877371788025 + ], + "credit_map": [ + 0.010744870640337467, + 0.007473092991858721, + 0.019774053245782852, + 0.009151649661362171, + 0.012196952477097511, + 0.011455941013991833, + 0.009477476589381695, + 0.016309762373566628, + 0.006656182464212179, + 0.00389373698271811, + 0.01381918229162693, + 0.020200802013278008, + 0.004695103969424963, + 0.0028284697327762842, + 0.006117283832281828, + 0.0030516954138875008, + 0.008290473371744156, + 0.01257288083434105, + 0.01830167882144451, + 0.018740499392151833, + 0.009159688837826252, + 0.0053868526592850685, + 0.008126205764710903, + 0.015551593154668808, + 0.006988650653511286, + 0.009821521118283272, + 0.01024347823113203, + 0.03944201394915581, + 0.009218443185091019, + 0.0035289176739752293, + 0.013428492471575737, + 0.010075608268380165, + 0.011133318766951561, + 0.007223037537187338, + 0.027737997472286224, + 0.017424320802092552, + 0.009228917770087719, + 0.008696767501533031, + 0.027115046977996826, + 0.025771088898181915, + 0.008283650502562523, + 0.0167987197637558, + 0.011628893204033375, + 0.008549993857741356, + 0.013299992308020592, + 0.015272303484380245, + 0.007582569494843483, + 0.006301140878349543, + 0.008265454322099686, + 0.004385017789900303, + 0.004060676787048578, + 0.007145005278289318, + 0.00662120571359992, + 0.007334383670240641, + 0.006134178955107927, + 0.003788864705711603, + 0.009658338502049446, + 0.008246311917901039, + 0.019818244501948357, + 0.011573156341910362, + 0.01663650944828987, + 0.012528356164693832, + 0.011297210119664669, + 0.017067456617951393, + 0.004693410359323025, + 0.005573610309511423, + 0.009802929125726223, + 0.0038343064952641726, + 0.005156394559890032, + 0.004902812652289867, + 0.00430037546902895, + 0.007417056709527969, + 0.025856688618659973, + 0.018839523196220398, + 0.07511822134256363, + 0.008921285159885883, + 0.01787431165575981, + 0.018087463453412056, + 0.007830166257917881, + 0.04846009239554405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0022025867365300655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633404780179262 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633404780179262, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633404780179262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08901868015527725, + 0.03736741840839386, + 0.07080764323472977, + 0.04945771023631096, + 0.04026324301958084, + 0.025541627779603004, + 0.03808961436152458, + 0.10125882923603058, + 0.08633751422166824, + 0.46185770630836487 + ], + "credit_map": [ + 0.01373843103647232, + 0.008200569078326225, + 0.018094506114721298, + 0.008532612584531307, + 0.008559733629226685, + 0.010877050459384918, + 0.009755539707839489, + 0.01126023568212986, + 0.00475596496835351, + 0.004514760337769985, + 0.009347422979772091, + 0.0027220877818763256, + 0.005801434628665447, + 0.0031702909618616104, + 0.0032277023419737816, + 0.0038277527783066034, + 0.006240594666451216, + 0.007526376284658909, + 0.00971142016351223, + 0.018704768270254135, + 0.0056564235128462315, + 0.003992789890617132, + 0.006172539666295052, + 0.012802733108401299, + 0.008164163678884506, + 0.005344776436686516, + 0.0062697636894881725, + 0.01121373102068901, + 0.006530236452817917, + 0.003616585396230221, + 0.005083377938717604, + 0.0032350770197808743, + 0.004012987017631531, + 0.002170899650081992, + 0.009742360562086105, + 0.004086238332092762, + 0.0034405640326440334, + 0.0031318413093686104, + 0.006264492403715849, + 0.007413857150822878, + 0.0023712394759058952, + 0.004353624302893877, + 0.003823252161964774, + 0.002755439840257168, + 0.0037655849009752274, + 0.0033553035464137793, + 0.002678966848179698, + 0.0024382176343351603, + 0.0032718142028898, + 0.0038066175766289234, + 0.0048095472157001495, + 0.004241174552589655, + 0.007548128254711628, + 0.006624211091548204, + 0.004377533216029406, + 0.0034105863887816668, + 0.008438317105174065, + 0.010063976049423218, + 0.020236877724528313, + 0.012396386824548244, + 0.01004765834659338, + 0.012668992392718792, + 0.011357403360307217, + 0.01604921743273735, + 0.0127748753875494, + 0.018869731575250626, + 0.008473836816847324, + 0.006267714314162731, + 0.011128201149404049, + 0.012589216232299805, + 0.008741634897887707, + 0.007492311764508486, + 0.05116332694888115, + 0.0424501895904541, + 0.1448986530303955, + 0.02458789385855198, + 0.039774734526872635, + 0.040655214339494705, + 0.02049863524734974, + 0.09782904386520386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004633404780179262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, full-length shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left in an elegant circular trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dancer wearing a flowing, gauzy silk Hanfu in pale celadon green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the courtyard framed by traditional Chinese architecture with dark wooden pillars and gray-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden light of late afternoon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901165425777435 + ], + "opposite_logprob_tokens": [ + -2.4640116691589355 + ], + "logprob_reward": -0.08901165425777435, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901165425777435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09792634844779968, + 0.06137044355273247, + 0.07204145938158035, + 0.09895815700292587, + 0.1211591362953186, + 0.03949199244379997, + 0.0958910584449768, + 0.043489959090948105, + 0.09202811866998672, + 0.2776432931423187 + ], + "credit_map": [ + 0.009597724303603172, + 0.007732874248176813, + 0.01723022200167179, + 0.012098009698092937, + 0.0178940761834383, + 0.011268246918916702, + 0.012330243363976479, + 0.009774954989552498, + 0.0077052670530974865, + 0.00642919447273016, + 0.004157686606049538, + 0.015322382561862469, + 0.007063338067382574, + 0.006429929286241531, + 0.006635722704231739, + 0.0076269241981208324, + 0.008599627763032913, + 0.007987765595316887, + 0.012025942094624043, + 0.014447900466620922, + 0.006442910525947809, + 0.004365548025816679, + 0.00664410088211298, + 0.011527659371495247, + 0.007412312086671591, + 0.00864277221262455, + 0.01203511469066143, + 0.032165057957172394, + 0.011666272766888142, + 0.004992642905563116, + 0.013010301627218723, + 0.009033690206706524, + 0.013493751175701618, + 0.005553120747208595, + 0.031668521463871, + 0.01264149323105812, + 0.007606787607073784, + 0.006259556859731674, + 0.020493507385253906, + 0.02344239503145218, + 0.004143351223319769, + 0.006007369142025709, + 0.004401613026857376, + 0.004636887926608324, + 0.006703191436827183, + 0.005957659333944321, + 0.00457609910517931, + 0.003065819386392832, + 0.01177443191409111, + 0.01190571766346693, + 0.009009317494928837, + 0.012481124140322208, + 0.01819353550672531, + 0.01552668958902359, + 0.010405139997601509, + 0.006595103535801172, + 0.006123528350144625, + 0.004981629084795713, + 0.005706609692424536, + 0.00410335510969162, + 0.004859797656536102, + 0.005007605534046888, + 0.0055401017889380455, + 0.007167333271354437, + 0.009427065961062908, + 0.009056787006556988, + 0.028490018099546432, + 0.007204206194728613, + 0.00998641736805439, + 0.006744961719959974, + 0.006454481743276119, + 0.01466417871415615, + 0.03131192922592163, + 0.019541194662451744, + 0.09211446344852448, + 0.014308808371424675, + 0.020723093301057816, + 0.027439920231699944, + 0.011119016446173191, + 0.061084866523742676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.08901165425777435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_tensor([6470, 6470])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_tensor([6470, 6470])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b3654a32456a053b70396cc5996281a3f57b486f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_tensor([6470, 6470])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, high-speed tracking shot, viewed from a slightly upward angle, closely follows a streamlined, futuristic spacecraft as it hurtles through the digital sky at breakneck speed. The vessel's hull, composed of a polished nanocomposite material resembling obsidian, is seamless and smooth, perfectly reflecting the light of the digital world below. Thin, cyan-blue energy lines flow along its aerodynamic contours, and its flanks are dotted with tiny, bright, pulsing navigation lights. The massive engines at the ship's rear belch blast hot blue plasma, creating a steady, dazzling trail of light. The force of this powerful thrust even slightly distorts the surrounding space. Beneath the ship, a vast, digital nightscape of the Guangdong-Hong Kong-Macao Greater Bay Area unfolds. This is not a physical city lightscape, but a three-dimensional holographic map composed of countless luminous blue and gold data streams. These data streams, like rushing rivers and highways, weave and pulsate through the virtual city skyline. As the spacecraft effortlessly zipped past at high speed, its trajectory created a fleeting ripple in these rivers of light. Digital particles, disturbed by its powerful aura, splattered to either side. Captured by a high-speed camera, the entire scene exhibited a breathtaking sense of dynamic beauty. The bright blue of the spacecraft's engines contrasted with the bluish-gold glow of the background data streams, creating a grand, futuristic atmosphere of high technology, extreme speed, and breathtaking speed.", + "seed": 18968, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_6470_seed_18968_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03091260143555701, + "vlm_logprob": -0.03091260143555701, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.03091260143555701, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013464237563312054 + ], + "opposite_logprob_tokens": [ + -4.388464450836182 + ], + "logprob_reward": -0.013464237563312054, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013464237563312054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26534947752952576, + 0.09869121760129929, + 0.06365292519330978, + 0.08206698298454285, + 0.059157054871320724, + 0.031525421887636185, + 0.05298769846558571, + 0.03320895880460739, + 0.03691745921969414, + 0.27644285559654236 + ], + "credit_map": [ + 0.02314094640314579, + 0.020103154703974724, + 0.03477711230516434, + 0.06054108217358589, + 0.04871819168329239, + 0.018822384998202324, + 0.03039180301129818, + 0.02885478362441063, + 0.01535315066576004, + 0.014936533756554127, + 0.015214753337204456, + 0.014319473877549171, + 0.009898288175463676, + 0.011476139537990093, + 0.009107491932809353, + 0.008385381661355495, + 0.0076703280210494995, + 0.011068630963563919, + 0.006889634765684605, + 0.008782087825238705, + 0.005675227381289005, + 0.010710181668400764, + 0.00655078561976552, + 0.0063060494139790535, + 0.009808261878788471, + 0.011584125459194183, + 0.007278543431311846, + 0.011441996321082115, + 0.010761158540844917, + 0.011479802429676056, + 0.009063364937901497, + 0.01064973697066307, + 0.006616681348532438, + 0.0073043727315962315, + 0.005973460152745247, + 0.007575934287160635, + 0.005911373067647219, + 0.013020815327763557, + 0.006876942701637745, + 0.005877472925931215, + 0.00341487443074584, + 0.004063422791659832, + 0.0013406160287559032, + 0.004879869055002928, + 0.004510597791522741, + 0.0038776255678385496, + 0.005552822723984718, + 0.0038855918683111668, + 0.007045856676995754, + 0.005463344044983387, + 0.005169641692191362, + 0.0032109220046550035, + 0.004294824320822954, + 0.006509795784950256, + 0.007540677674114704, + 0.01375263649970293, + 0.005167811643332243, + 0.00393626419827342, + 0.0026589143089950085, + 0.004447535146027803, + 0.003377471351996064, + 0.0064229657873511314, + 0.0034221189562231302, + 0.003775876248255372, + 0.0032500203233212233, + 0.004018689505755901, + 0.003044029464945197, + 0.005697564221918583, + 0.004393734969198704, + 0.006104131694883108, + 0.0048818993382155895, + 0.0055273897014558315, + 0.03239182382822037, + 0.0230785571038723, + 0.022438660264015198, + 0.01826435886323452, + 0.020880311727523804, + 0.06905689835548401, + 0.058930277824401855, + 0.03140196576714516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.013464237563312054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028368488419800997 + ], + "opposite_logprob_tokens": [ + -5.87783670425415 + ], + "logprob_reward": -0.0028368488419800997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028368488419800997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09463346749544144, + 0.06463470309972763, + 0.08375227451324463, + 0.1060221940279007, + 0.1563301533460617, + 0.08643540740013123, + 0.08023196458816528, + 0.05387085676193237, + 0.038260605186223984, + 0.23582839965820312 + ], + "credit_map": [ + 0.010295979678630829, + 0.007896420545876026, + 0.010859573259949684, + 0.015633611008524895, + 0.016195112839341164, + 0.00755283422768116, + 0.011981005780398846, + 0.014218919910490513, + 0.011584949679672718, + 0.008582071401178837, + 0.008343013003468513, + 0.009593421593308449, + 0.007980444468557835, + 0.0063254693523049355, + 0.006829431280493736, + 0.005395899061113596, + 0.00972567219287157, + 0.009624296799302101, + 0.010814170353114605, + 0.015155663713812828, + 0.010536005720496178, + 0.010184230282902718, + 0.010569833219051361, + 0.007142392918467522, + 0.015606964007019997, + 0.012399386614561081, + 0.011471468955278397, + 0.014469821006059647, + 0.01485027838498354, + 0.015317958779633045, + 0.011658795177936554, + 0.010247503407299519, + 0.02371126040816307, + 0.019986283034086227, + 0.02046070247888565, + 0.015827415511012077, + 0.019055623561143875, + 0.020817531272768974, + 0.01964535377919674, + 0.01682596281170845, + 0.006579461041837931, + 0.010771527886390686, + 0.0032978758681565523, + 0.01152349915355444, + 0.013424335047602654, + 0.011836579069495201, + 0.018476972356438637, + 0.010525154881179333, + 0.009980863891541958, + 0.009530982933938503, + 0.005256256554275751, + 0.006536536384373903, + 0.00847046822309494, + 0.009358866140246391, + 0.012216882780194283, + 0.018881095573306084, + 0.007083911914378405, + 0.005892153829336166, + 0.004121941979974508, + 0.0037757637910544872, + 0.006683060899376869, + 0.00798984244465828, + 0.0050018769688904285, + 0.013322299346327782, + 0.005445878021419048, + 0.004348542541265488, + 0.004054203629493713, + 0.005329248961061239, + 0.004400281235575676, + 0.004427746869623661, + 0.004004243761301041, + 0.006250455044209957, + 0.028677260503172874, + 0.018462128937244415, + 0.018823517486453056, + 0.01460415031760931, + 0.018730314448475838, + 0.06150248274207115, + 0.04984261840581894, + 0.025185931473970413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.0028368488419800997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18450862169265747, + 0.10971038043498993, + 0.1802823543548584, + 0.05378669872879982, + 0.03834909945726395, + 0.037532709538936615, + 0.054298050701618195, + 0.040424421429634094, + 0.04759801924228668, + 0.25350967049598694 + ], + "credit_map": [ + 0.01628389209508896, + 0.014555182307958603, + 0.027584997937083244, + 0.0411115437746048, + 0.03006180375814438, + 0.014761618338525295, + 0.018373748287558556, + 0.02177584171295166, + 0.024766460061073303, + 0.01679392158985138, + 0.013220824301242828, + 0.010347473435103893, + 0.011768968775868416, + 0.009990972466766834, + 0.011659398674964905, + 0.01116235926747322, + 0.020695695653557777, + 0.01761535555124283, + 0.024823907762765884, + 0.029149996116757393, + 0.024743786081671715, + 0.022896135225892067, + 0.025670770555734634, + 0.014686697162687778, + 0.0056523834355175495, + 0.009586603380739689, + 0.003934669308364391, + 0.006903855595737696, + 0.005954930558800697, + 0.006786650512367487, + 0.006099834572523832, + 0.008867770433425903, + 0.004120047204196453, + 0.0040483674965798855, + 0.003376607783138752, + 0.004892571363598108, + 0.0057620517909526825, + 0.007299911696463823, + 0.005145745817571878, + 0.0037037977017462254, + 0.004066128749400377, + 0.004517342429608107, + 0.0016554618487134576, + 0.004540346562862396, + 0.005660259630531073, + 0.004586807452142239, + 0.008057461120188236, + 0.00444890046492219, + 0.0075979954563081264, + 0.006399249192327261, + 0.005705575924366713, + 0.0036093073431402445, + 0.004505175165832043, + 0.005662320181727409, + 0.006944195833057165, + 0.013874229043722153, + 0.005544429179280996, + 0.003775273682549596, + 0.003161604516208172, + 0.0036000320687890053, + 0.00563735282048583, + 0.009365859441459179, + 0.0033935054671019316, + 0.005946365650743246, + 0.005268642213195562, + 0.005488344002515078, + 0.003912216052412987, + 0.007113609462976456, + 0.004757250659167767, + 0.006303438451141119, + 0.006579704117029905, + 0.008174817077815533, + 0.030644303187727928, + 0.020256781950592995, + 0.022205691784620285, + 0.017508791759610176, + 0.020750468596816063, + 0.05956625938415527, + 0.053047750145196915, + 0.0295296348631382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.00525581045076251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270591914653778 + ], + "opposite_logprob_tokens": [ + -2.1270592212677 + ], + "logprob_reward": -0.1270591914653778, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270591914653778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07532122731208801, + 0.05415508151054382, + 0.03660460561513901, + 0.062365975230932236, + 0.08561766147613525, + 0.039272479712963104, + 0.030893493443727493, + 0.06293078511953354, + 0.05687597021460533, + 0.4959627389907837 + ], + "credit_map": [ + 0.007420704234391451, + 0.006444623693823814, + 0.009183250367641449, + 0.016118772327899933, + 0.010991610586643219, + 0.005670943297445774, + 0.00841639656573534, + 0.01107492484152317, + 0.00935336109250784, + 0.007343371864408255, + 0.0067020901478827, + 0.005531949456781149, + 0.005978010129183531, + 0.005345708690583706, + 0.00621882313862443, + 0.0076817711815238, + 0.004373436793684959, + 0.00415883818641305, + 0.0044761416502296925, + 0.007158762775361538, + 0.003358318004757166, + 0.006447939667850733, + 0.00351019948720932, + 0.0031209702137857676, + 0.008188349194824696, + 0.008860823698341846, + 0.004920240491628647, + 0.008953534997999668, + 0.00836693774908781, + 0.009218502789735794, + 0.0068069021217525005, + 0.007050681859254837, + 0.008467097766697407, + 0.007031777873635292, + 0.01521389465779066, + 0.007992028258740902, + 0.012094162404537201, + 0.02180846780538559, + 0.007069064769893885, + 0.0059411716647446156, + 0.004422585014253855, + 0.005148079711943865, + 0.0016896738670766354, + 0.003996755462139845, + 0.006249483209103346, + 0.004819985944777727, + 0.008845572359859943, + 0.004100344143807888, + 0.004597074817866087, + 0.004586127121001482, + 0.0033236746676266193, + 0.002254907041788101, + 0.004771618638187647, + 0.002866210648790002, + 0.004444661550223827, + 0.004049220122396946, + 0.007334111724048853, + 0.005404291208833456, + 0.003996056038886309, + 0.007283971644937992, + 0.005612790584564209, + 0.01988058164715767, + 0.0053967698477208614, + 0.008022214286029339, + 0.0067979274317622185, + 0.007117682136595249, + 0.005408658646047115, + 0.009926406666636467, + 0.007011513225734234, + 0.0067512718960642815, + 0.007588203065097332, + 0.006274308543652296, + 0.06449131667613983, + 0.03868725895881653, + 0.04492153599858284, + 0.03347425535321236, + 0.039341893047094345, + 0.0898120179772377, + 0.11159680038690567, + 0.0736376941204071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.1270591914653778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946918856352568 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005946918856352568, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946918856352568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21323299407958984, + 0.10902299731969833, + 0.07255321741104126, + 0.08031900972127914, + 0.07344097644090652, + 0.09035992622375488, + 0.1407172530889511, + 0.05947896093130112, + 0.04018879681825638, + 0.12068580836057663 + ], + "credit_map": [ + 0.01683906652033329, + 0.021506909281015396, + 0.03955313563346863, + 0.05411753058433533, + 0.02441287226974964, + 0.014646541327238083, + 0.022220268845558167, + 0.019936664029955864, + 0.019110407680273056, + 0.018282538279891014, + 0.014946228824555874, + 0.01175912469625473, + 0.010826683603227139, + 0.009618899784982204, + 0.013662073761224747, + 0.010817048139870167, + 0.008646841160953045, + 0.008073793724179268, + 0.00927857868373394, + 0.009976185858249664, + 0.0077222115360200405, + 0.011352213099598885, + 0.009999622590839863, + 0.0075037782080471516, + 0.010658130049705505, + 0.014965378679335117, + 0.005700398236513138, + 0.00509211141616106, + 0.007059501018375158, + 0.018603527918457985, + 0.008897819556295872, + 0.009342141449451447, + 0.014961214736104012, + 0.012584404088556767, + 0.010129527188837528, + 0.006518393289297819, + 0.00825449638068676, + 0.007997424341738224, + 0.007210204843431711, + 0.005785306915640831, + 0.00735607510432601, + 0.01504139881581068, + 0.0035431438591331244, + 0.012538359500467777, + 0.013008403591811657, + 0.011276330798864365, + 0.016620388254523277, + 0.010975820943713188, + 0.017602108418941498, + 0.015972815454006195, + 0.010402756743133068, + 0.010507461614906788, + 0.012806907296180725, + 0.01548662781715393, + 0.022013124078512192, + 0.035925451666116714, + 0.004906749818474054, + 0.004207357298582792, + 0.0037328831385821104, + 0.01131918653845787, + 0.005816943943500519, + 0.0162724107503891, + 0.004588984418660402, + 0.008634441532194614, + 0.004642151761800051, + 0.004219589754939079, + 0.0029282779432833195, + 0.0039533693343400955, + 0.00431471923366189, + 0.005605035927146673, + 0.004775049164891243, + 0.009750603698194027, + 0.01894981600344181, + 0.010933341458439827, + 0.010343474335968494, + 0.00928647629916668, + 0.009534274227917194, + 0.022795725613832474, + 0.024035271257162094, + 0.014807438477873802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.005946918856352568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, high-speed tracking shot, viewed from a slightly upward angle, closely follows a streamlined, futuristic spacecraft as it hurtles through the digital sky at breakneck speed. The vessel's hull, composed of a polished nanocomposite material resembling obsidian, is seamless and smooth, perfectly reflecting the light of the digital world below. Thin, cyan-blue energy lines flow along its aerodynamic contours, and its flanks are dotted with tiny, bright, pulsing navigation lights. The massive engines at the ship's rear belch blast hot blue plasma, creating a steady, dazzling trail of light. The force of this powerful thrust even slightly distorts the surrounding space. Beneath the ship, a vast, digital nightscape of the Guangdong-Hong Kong-Macao Greater Bay Area unfolds. This is not a physical city lightscape, but a three-dimensional holographic map composed of countless luminous blue and gold data streams. These data streams, like rushing rivers and highways, weave and pulsate through the virtual city skyline. As the spacecraft effortlessly zipped past at high speed, its trajectory created a fleeting ripple in these rivers of light. Digital particles, disturbed by its powerful aura, splattered to either side. Captured by a high-speed camera, the entire scene exhibited a breathtaking sense of dynamic beauty. The bright blue of the spacecraft's engines contrasted with the bluish-gold glow of the background data streams, creating a grand, futuristic atmosphere of high technology, extreme speed, and breathtaking speed.", + "seed": 18969, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_6470_seed_18969_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022940581245347858, + "vlm_logprob": -0.022940581245347858, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.022940581245347858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017253074795007706 + ], + "opposite_logprob_tokens": [ + -4.1422529220581055 + ], + "logprob_reward": -0.017253074795007706, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017253074795007706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2713192105293274, + 0.13558243215084076, + 0.16998374462127686, + 0.04742734134197235, + 0.09702528268098831, + 0.06736470758914948, + 0.03646717593073845, + 0.05285792052745819, + 0.05185762420296669, + 0.07011450082063675 + ], + "credit_map": [ + 0.020324477925896645, + 0.024624856188893318, + 0.050496816635131836, + 0.07262516766786575, + 0.02611500211060047, + 0.025919677689671516, + 0.025449106469750404, + 0.025764090940356255, + 0.01340471114963293, + 0.00962775107473135, + 0.015858979895710945, + 0.016492366790771484, + 0.01241315994411707, + 0.02889818325638771, + 0.018418602645397186, + 0.02046867646276951, + 0.019774511456489563, + 0.03137955814599991, + 0.01050479058176279, + 0.012685578316450119, + 0.03892785310745239, + 0.021272074431180954, + 0.019401712343096733, + 0.01603766158223152, + 0.010750290006399155, + 0.00615238631144166, + 0.0026330493856221437, + 0.005428512580692768, + 0.004244999028742313, + 0.005909225437790155, + 0.004405261017382145, + 0.00790361873805523, + 0.009986773133277893, + 0.02027272991836071, + 0.007599370554089546, + 0.00996728241443634, + 0.008089293725788593, + 0.01042091939598322, + 0.0196752417832613, + 0.01101367175579071, + 0.007638601120561361, + 0.010084723122417927, + 0.005046806298196316, + 0.006371047347784042, + 0.0076100025326013565, + 0.007980970665812492, + 0.011208660900592804, + 0.011423898860812187, + 0.0047523146495223045, + 0.002428415697067976, + 0.005551515147089958, + 0.0031052129343152046, + 0.005161313805729151, + 0.0061197951436042786, + 0.004558729939162731, + 0.0047898790799081326, + 0.004551892634481192, + 0.005009165965020657, + 0.006875744089484215, + 0.008476275019347668, + 0.004241142421960831, + 0.007579098455607891, + 0.006707547232508659, + 0.009417052380740643, + 0.00485201645642519, + 0.007925201207399368, + 0.010190119966864586, + 0.005734968930482864, + 0.0051302979700267315, + 0.00434405542910099, + 0.007565206382423639, + 0.006115758791565895, + 0.003498959355056286, + 0.006296006962656975, + 0.011224900372326374, + 0.012846981175243855, + 0.0109552051872015, + 0.009320617653429508, + 0.007817075587809086, + 0.00815475545823574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.017253074795007706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003211818402633071 + ], + "opposite_logprob_tokens": [ + -5.753211975097656 + ], + "logprob_reward": -0.003211818402633071, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003211818402633071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3133579194545746, + 0.12852296233177185, + 0.07942979782819748, + 0.0951385349035263, + 0.0517718642950058, + 0.06586340069770813, + 0.0743301659822464, + 0.05049464479088783, + 0.04435472935438156, + 0.09673596918582916 + ], + "credit_map": [ + 0.022098112851381302, + 0.021067669615149498, + 0.06653950363397598, + 0.10509850829839706, + 0.019521882757544518, + 0.024004381150007248, + 0.025916539132595062, + 0.029111335054039955, + 0.01349487341940403, + 0.009058604016900063, + 0.016006311401724815, + 0.022352013736963272, + 0.009182077832520008, + 0.02263914979994297, + 0.016273140907287598, + 0.019516799598932266, + 0.00879704486578703, + 0.014561358839273453, + 0.006580693647265434, + 0.006283341906964779, + 0.018998662009835243, + 0.010042798705399036, + 0.006869425531476736, + 0.007296474184840918, + 0.017273303121328354, + 0.011007335968315601, + 0.006271385587751865, + 0.01094850618392229, + 0.01065643597394228, + 0.013942611403763294, + 0.010800272226333618, + 0.0142386918887496, + 0.004862479865550995, + 0.006044881884008646, + 0.0029843533411622047, + 0.005202918313443661, + 0.004357791040092707, + 0.009829090908169746, + 0.011953809298574924, + 0.006536541040986776, + 0.004481456708163023, + 0.0067071812227368355, + 0.004632178228348494, + 0.004325359594076872, + 0.004489121492952108, + 0.008400553837418556, + 0.01635756716132164, + 0.016469981521368027, + 0.00671594450250268, + 0.005892026238143444, + 0.011116738431155682, + 0.0074867624789476395, + 0.01139586791396141, + 0.016161678358912468, + 0.00865785963833332, + 0.006903282832354307, + 0.005225343629717827, + 0.004115000367164612, + 0.006455355789512396, + 0.007803068961948156, + 0.0051822662353515625, + 0.007914499379694462, + 0.006468082778155804, + 0.007331027649343014, + 0.004188072867691517, + 0.0052373274229466915, + 0.008937927894294262, + 0.005499574821442366, + 0.004568879492580891, + 0.004837507382035255, + 0.005396244116127491, + 0.005689193494617939, + 0.0053927358239889145, + 0.008629418909549713, + 0.013691646978259087, + 0.015570685267448425, + 0.016301767900586128, + 0.013362321071326733, + 0.011399463750422001, + 0.012387925758957863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003211818402633071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.250518798828125, + 0.11183396726846695, + 0.1398552656173706, + 0.05430029332637787, + 0.039198726415634155, + 0.07053305953741074, + 0.14068050682544708, + 0.058604829013347626, + 0.0705181360244751, + 0.06395646184682846 + ], + "credit_map": [ + 0.020186837762594223, + 0.016702724620699883, + 0.0479922816157341, + 0.0703580230474472, + 0.02377515472471714, + 0.02079811878502369, + 0.027828291058540344, + 0.022877370938658714, + 0.011302010156214237, + 0.007339757401496172, + 0.011883514001965523, + 0.012577607296407223, + 0.009341123513877392, + 0.02672242559492588, + 0.016496043652296066, + 0.01617148518562317, + 0.0171342883259058, + 0.025636723265051842, + 0.008698529563844204, + 0.009572392329573631, + 0.03323206305503845, + 0.017980193719267845, + 0.015247892588376999, + 0.012353192083537579, + 0.010067719034850597, + 0.006261403206735849, + 0.00408199755474925, + 0.005695092026144266, + 0.005960931070148945, + 0.006626761052757502, + 0.0062635475769639015, + 0.009342839941382408, + 0.00658206082880497, + 0.006799993105232716, + 0.0037240376695990562, + 0.003806081600487232, + 0.0027879655826836824, + 0.005589955486357212, + 0.005626655649393797, + 0.004281976725906134, + 0.007019507698714733, + 0.007699819281697273, + 0.005399220623075962, + 0.0048827119171619415, + 0.0049903010949492455, + 0.009312105365097523, + 0.016471177339553833, + 0.014758223667740822, + 0.012308835051953793, + 0.00715669197961688, + 0.019380878657102585, + 0.01442668866366148, + 0.02460474707186222, + 0.02521353028714657, + 0.02292320877313614, + 0.01466592587530613, + 0.005544556770473719, + 0.005018402822315693, + 0.007434780243784189, + 0.009582224301993847, + 0.005347183905541897, + 0.008592084981501102, + 0.0072471341118216515, + 0.009838459081947803, + 0.004475158639252186, + 0.0075981090776622295, + 0.014678997918963432, + 0.00994102843105793, + 0.005478686187416315, + 0.008280151523649693, + 0.011500909924507141, + 0.008565095253288746, + 0.0033585287164896727, + 0.005268002860248089, + 0.010555755347013474, + 0.012612087652087212, + 0.009980255737900734, + 0.008412612602114677, + 0.0074248602613806725, + 0.006344358902424574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897964119911194 + ], + "opposite_logprob_tokens": [ + -2.578979730606079 + ], + "logprob_reward": -0.07897964119911194, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897964119911194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29868143796920776, + 0.13897307217121124, + 0.0523485504090786, + 0.048555634915828705, + 0.039536118507385254, + 0.1288367360830307, + 0.06779991090297699, + 0.08836782723665237, + 0.058775972574949265, + 0.07812471687793732 + ], + "credit_map": [ + 0.031040417030453682, + 0.03743797913193703, + 0.040399763733148575, + 0.04591814801096916, + 0.03497153893113136, + 0.0340237133204937, + 0.03793177381157875, + 0.036958079785108566, + 0.015469101257622242, + 0.010008726269006729, + 0.014358023181557655, + 0.015967123210430145, + 0.011207597330212593, + 0.031031355261802673, + 0.01906842738389969, + 0.02186271734535694, + 0.0060617392882704735, + 0.007820572704076767, + 0.003302036551758647, + 0.0035374064464122057, + 0.011617480777204037, + 0.006994729861617088, + 0.006184544414281845, + 0.006830041762441397, + 0.00909124594181776, + 0.005342559888958931, + 0.0037229799199849367, + 0.004369261208921671, + 0.005287745967507362, + 0.0073043713346123695, + 0.00606511952355504, + 0.007372352294623852, + 0.0044587342999875546, + 0.006957939825952053, + 0.002961505204439163, + 0.0045538414269685745, + 0.0030516176484525204, + 0.004879486747086048, + 0.008152535185217857, + 0.004520459100604057, + 0.010428005829453468, + 0.01041608676314354, + 0.008378459140658379, + 0.007149998564273119, + 0.01004064828157425, + 0.017710037529468536, + 0.032396137714385986, + 0.03231734782457352, + 0.004898556973785162, + 0.003192777745425701, + 0.010836041532456875, + 0.004379485733807087, + 0.007773147895932198, + 0.019535761326551437, + 0.00940354447811842, + 0.007780596613883972, + 0.008072933182120323, + 0.010527962818741798, + 0.012882026843726635, + 0.011936481110751629, + 0.007358008995652199, + 0.010582363232970238, + 0.009100416675209999, + 0.017907639965415, + 0.005176275037229061, + 0.006774245761334896, + 0.009765722788870335, + 0.010235526598989964, + 0.005883835721760988, + 0.006191903725266457, + 0.006350717507302761, + 0.00839774589985609, + 0.004507143050432205, + 0.006588257383555174, + 0.011702185496687889, + 0.01236442569643259, + 0.014749430119991302, + 0.009577260352671146, + 0.008365119807422161, + 0.010270893573760986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07897964119911194 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626938167959452 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007626938167959452, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626938167959452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45475730299949646, + 0.08329013735055923, + 0.05921011418104172, + 0.03777638450264931, + 0.04492231458425522, + 0.04090793430805206, + 0.102919802069664, + 0.043466098606586456, + 0.06801707297563553, + 0.06473281979560852 + ], + "credit_map": [ + 0.044778741896152496, + 0.03712594881653786, + 0.07786544412374496, + 0.11125528067350388, + 0.03506119176745415, + 0.046282585710287094, + 0.06931184977293015, + 0.03307627886533737, + 0.009436214342713356, + 0.00588248111307621, + 0.008477270603179932, + 0.010341137647628784, + 0.005686039570719004, + 0.017847923561930656, + 0.014262516051530838, + 0.011356550268828869, + 0.006741546560078859, + 0.007400970906019211, + 0.005867870524525642, + 0.006328035145998001, + 0.01137478370219469, + 0.008410956710577011, + 0.006193295121192932, + 0.006892655044794083, + 0.004415613133460283, + 0.0035026902332901955, + 0.002562141278758645, + 0.004078201949596405, + 0.004130159970372915, + 0.005082000978291035, + 0.005395873915404081, + 0.008609704673290253, + 0.006401035934686661, + 0.005975045263767242, + 0.0026013036258518696, + 0.004902889020740986, + 0.0036223421338945627, + 0.006909833289682865, + 0.009382423013448715, + 0.0051274425350129604, + 0.0055907657369971275, + 0.005340580828487873, + 0.00416021840646863, + 0.003245533676818013, + 0.0042338380590081215, + 0.0049864803440868855, + 0.007150487508624792, + 0.006200031843036413, + 0.006748996675014496, + 0.0036269829142838717, + 0.016334621235728264, + 0.00634039519354701, + 0.010974201373755932, + 0.03254975005984306, + 0.014815576374530792, + 0.0115292863920331, + 0.004417033866047859, + 0.0040062665939331055, + 0.008017142303287983, + 0.005693074315786362, + 0.0030847834423184395, + 0.00686030276119709, + 0.004413549788296223, + 0.006973945535719395, + 0.005497802514582872, + 0.00946546159684658, + 0.013542760163545609, + 0.008932603523135185, + 0.006377811543643475, + 0.005928054451942444, + 0.009202965535223484, + 0.009069612249732018, + 0.003682613605633378, + 0.00554570322856307, + 0.010066901333630085, + 0.01120294351130724, + 0.010436587035655975, + 0.010563770309090614, + 0.00679335743188858, + 0.006440946366637945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.007626938167959452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, high-speed tracking shot, viewed from a slightly upward angle, closely follows a streamlined, futuristic spacecraft as it hurtles through the digital sky at breakneck speed. The vessel's hull, composed of a polished nanocomposite material resembling obsidian, is seamless and smooth, perfectly reflecting the light of the digital world below. Thin, cyan-blue energy lines flow along its aerodynamic contours, and its flanks are dotted with tiny, bright, pulsing navigation lights. The massive engines at the ship's rear belch blast hot blue plasma, creating a steady, dazzling trail of light. The force of this powerful thrust even slightly distorts the surrounding space. Beneath the ship, a vast, digital nightscape of the Guangdong-Hong Kong-Macao Greater Bay Area unfolds. This is not a physical city lightscape, but a three-dimensional holographic map composed of countless luminous blue and gold data streams. These data streams, like rushing rivers and highways, weave and pulsate through the virtual city skyline. As the spacecraft effortlessly zipped past at high speed, its trajectory created a fleeting ripple in these rivers of light. Digital particles, disturbed by its powerful aura, splattered to either side. Captured by a high-speed camera, the entire scene exhibited a breathtaking sense of dynamic beauty. The bright blue of the spacecraft's engines contrasted with the bluish-gold glow of the background data streams, creating a grand, futuristic atmosphere of high technology, extreme speed, and breathtaking speed.", + "seed": 18970, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_6470_seed_18970_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019360352447256444, + "vlm_logprob": -0.019360352447256444, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.019360352447256444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010829368606209755 + ], + "opposite_logprob_tokens": [ + -4.635829448699951 + ], + "logprob_reward": -0.010829368606209755, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010829368606209755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07563649117946625, + 0.04050692543387413, + 0.08696205168962479, + 0.02271469496190548, + 0.016830185428261757, + 0.019784539937973022, + 0.04026734083890915, + 0.03911864012479782, + 0.4930512011051178, + 0.16512790322303772 + ], + "credit_map": [ + 0.01253463514149189, + 0.009846615605056286, + 0.010868717916309834, + 0.011020662263035774, + 0.007727391552180052, + 0.006309977732598782, + 0.008786280639469624, + 0.008542211726307869, + 0.00844333041459322, + 0.005675943568348885, + 0.00845356471836567, + 0.004885394591838121, + 0.0026249438524246216, + 0.003432502970099449, + 0.003572231624275446, + 0.003419012762606144, + 0.012057141400873661, + 0.014519927091896534, + 0.011428236030042171, + 0.009839741513133049, + 0.008965060114860535, + 0.0074205463752150536, + 0.011029229499399662, + 0.011702173389494419, + 0.004220125265419483, + 0.0034643488470464945, + 0.002697966992855072, + 0.002709947293624282, + 0.0025854911655187607, + 0.002229992998763919, + 0.0025341680739074945, + 0.002272654790431261, + 0.0015175932785496116, + 0.003727107308804989, + 0.0023664766922593117, + 0.0014701655600219965, + 0.001599613344296813, + 0.0021794550120830536, + 0.0014017955400049686, + 0.002567978110164404, + 0.0022982966620475054, + 0.0019333693198859692, + 0.0024293153546750546, + 0.0022446634247899055, + 0.0024966797791421413, + 0.003820872399955988, + 0.002250818070024252, + 0.0023105230648070574, + 0.003700852394104004, + 0.0036710004787892103, + 0.0031194479670375586, + 0.0033568679355084896, + 0.013515613973140717, + 0.004760653246194124, + 0.0030465268064290285, + 0.005096377804875374, + 0.004220138769596815, + 0.0033715663012117147, + 0.002257969696074724, + 0.0026229091454297304, + 0.006658724043518305, + 0.008625754155218601, + 0.007440332323312759, + 0.003921241965144873, + 0.030378444120287895, + 0.0221576988697052, + 0.016993768513202667, + 0.01746046356856823, + 0.023446861654520035, + 0.03437395766377449, + 0.17016221582889557, + 0.178077831864357, + 0.008560741320252419, + 0.013436529785394669, + 0.03554108366370201, + 0.012442527338862419, + 0.06110893562436104, + 0.017287734895944595, + 0.010114089585840702, + 0.0066362665966153145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.010829368606209755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032116996590048075 + ], + "opposite_logprob_tokens": [ + -5.753211498260498 + ], + "logprob_reward": -0.0032116996590048075, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032116996590048075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025755342096090317, + 0.03591112419962883, + 0.014171715825796127, + 0.017936168238520622, + 0.043219782412052155, + 0.0245185699313879, + 0.08170418441295624, + 0.0577026903629303, + 0.5637351870536804, + 0.1353452205657959 + ], + "credit_map": [ + 0.005932601168751717, + 0.0037450087256729603, + 0.004693066235631704, + 0.0024394425563514233, + 0.002119589364156127, + 0.0017963757272809744, + 0.002319954801350832, + 0.002709304681047797, + 0.009354319423437119, + 0.005122554022818804, + 0.006496877875179052, + 0.003755174344405532, + 0.002237143227830529, + 0.0037387229967862368, + 0.0029420298524200916, + 0.0022642987314611673, + 0.0024610941763967276, + 0.002148838946595788, + 0.0020461094100028276, + 0.002798052504658699, + 0.0012259570648893714, + 0.0008529579499736428, + 0.001435953308828175, + 0.0012027524644508958, + 0.0025466661900281906, + 0.0025262932758778334, + 0.0023245743941515684, + 0.002030525356531143, + 0.0015487134223803878, + 0.0020192936062812805, + 0.002495448337867856, + 0.0024446535389870405, + 0.004854978993535042, + 0.006767821963876486, + 0.008369670249521732, + 0.003469428513199091, + 0.0042062923312187195, + 0.004850562661886215, + 0.0033413260243833065, + 0.007359698880463839, + 0.0024077112320810556, + 0.002910127630457282, + 0.002843917580321431, + 0.0026565578300505877, + 0.003457119222730398, + 0.004529909696429968, + 0.0031030410900712013, + 0.0026101854164153337, + 0.008626265451312065, + 0.009145180694758892, + 0.005661030765622854, + 0.006202544551342726, + 0.02382371947169304, + 0.011024968698620796, + 0.0065847886726260185, + 0.010635685175657272, + 0.007341952528804541, + 0.0071029504761099815, + 0.00371392909437418, + 0.004999994300305843, + 0.008438999764621258, + 0.012342226691544056, + 0.0076844291761517525, + 0.006078210659325123, + 0.03577888011932373, + 0.025478940457105637, + 0.019685113802552223, + 0.020304933190345764, + 0.027035236358642578, + 0.035505104809999466, + 0.19618742167949677, + 0.20375952124595642, + 0.0074881077744066715, + 0.007566459942609072, + 0.03151799365878105, + 0.02793002873659134, + 0.032606709748506546, + 0.012729865498840809, + 0.009349129162728786, + 0.0061569237150251865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0032116996590048075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642413653433323 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642413653433323, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642413653433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02845717780292034, + 0.018814368173480034, + 0.01087138056755066, + 0.009149322286248207, + 0.02360369823873043, + 0.010249895043671131, + 0.03803865239024162, + 0.040890879929065704, + 0.7364940643310547, + 0.08343064039945602 + ], + "credit_map": [ + 0.0053747352212667465, + 0.003420597640797496, + 0.004086860455572605, + 0.003043081844225526, + 0.0024587903171777725, + 0.004619978368282318, + 0.002951096510514617, + 0.0025020353496074677, + 0.004414823837578297, + 0.00307015934959054, + 0.0034005113411694765, + 0.0019476464949548244, + 0.0011913429480046034, + 0.0019010363612324, + 0.001369728590361774, + 0.0015191171551123261, + 0.0014333605067804456, + 0.0014406387927010655, + 0.001796383992768824, + 0.0023585110902786255, + 0.0009230056894011796, + 0.0008075156947597861, + 0.0011210921220481396, + 0.0009908712236210704, + 0.0013952386798337102, + 0.0013887719251215458, + 0.0009907784406095743, + 0.0008423785329796374, + 0.0008349172421731055, + 0.0011156921973451972, + 0.0009967926889657974, + 0.001584751415066421, + 0.0025400894228368998, + 0.003548696869984269, + 0.003165970789268613, + 0.001893608015961945, + 0.0025663936976343393, + 0.002972753718495369, + 0.002247926313430071, + 0.004668256267905235, + 0.0012521520256996155, + 0.0011956655653193593, + 0.0011848810827359557, + 0.0012039850698783994, + 0.0011134120868518949, + 0.0017044279957190156, + 0.001407518284395337, + 0.0011878523509949446, + 0.004074468743056059, + 0.003782691666856408, + 0.0032871330622583628, + 0.004684250336140394, + 0.008671618066728115, + 0.005790379364043474, + 0.003497539786621928, + 0.004250567406415939, + 0.005595976486802101, + 0.004039552994072437, + 0.0024426227901130915, + 0.003418281674385071, + 0.0050102644599974155, + 0.010517817921936512, + 0.005468321964144707, + 0.004398038145154715, + 0.04687744006514549, + 0.0340598002076149, + 0.02553124725818634, + 0.0263254102319479, + 0.035795532166957855, + 0.04293274134397507, + 0.2577671408653259, + 0.2672046422958374, + 0.0057640173472464085, + 0.007277177181094885, + 0.009917042218148708, + 0.009851117618381977, + 0.03186391294002533, + 0.007831796072423458, + 0.005585896782577038, + 0.005339676048606634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.008642413653433323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002337276935577 + ], + "opposite_logprob_tokens": [ + -2.69502329826355 + ], + "logprob_reward": -0.07002337276935577, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002337276935577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014150123111903667, + 0.006545893382281065, + 0.00488175917416811, + 0.004442078527063131, + 0.00634931493550539, + 0.004970148205757141, + 0.016218893229961395, + 0.057040583342313766, + 0.6159660816192627, + 0.2694351077079773 + ], + "credit_map": [ + 0.0028745445888489485, + 0.0021195639856159687, + 0.002885762369260192, + 0.0018992898985743523, + 0.0010988060384988785, + 0.0014196776319295168, + 0.0008972715004347265, + 0.0009552061674185097, + 0.0015473265666514635, + 0.0010464058723300695, + 0.0008650576928630471, + 0.0008639829466119409, + 0.0004319911531638354, + 0.0005735294544138014, + 0.0006615917081944644, + 0.0005560078425332904, + 0.0007230803021229804, + 0.0007963018724694848, + 0.0006839423440396786, + 0.0008171104127541184, + 0.00035433456650935113, + 0.0005054718931205571, + 0.00044602397247217596, + 0.0005554932868108153, + 0.0007604004931636155, + 0.0005561051075346768, + 0.000667847110889852, + 0.00038938564830459654, + 0.0003470816882327199, + 0.0005317293689586222, + 0.000530933088157326, + 0.0006585955270566046, + 0.0009150232654064894, + 0.0009693362517282367, + 0.0008784605306573212, + 0.0005047685117460787, + 0.0006410145433619618, + 0.0009074219851754606, + 0.000862155866343528, + 0.0006711337482556701, + 0.0005388951976783574, + 0.0005955143715254962, + 0.0007141935057006776, + 0.0004851915582548827, + 0.000575739424675703, + 0.0009182518115267158, + 0.000521525857038796, + 0.0006208362174220383, + 0.0018584985518828034, + 0.001785671105608344, + 0.0013394795823842287, + 0.001466844929382205, + 0.004658181220293045, + 0.0021448438055813313, + 0.0012254964094609022, + 0.00173987727612257, + 0.00806310772895813, + 0.007271704263985157, + 0.003563879756256938, + 0.005305663216859102, + 0.009095285087823868, + 0.010649127885699272, + 0.007593128830194473, + 0.0054986863397061825, + 0.03807468339800835, + 0.027378670871257782, + 0.021333809942007065, + 0.022149045020341873, + 0.029132820665836334, + 0.043921589851379395, + 0.2118368297815323, + 0.22213862836360931, + 0.013068396598100662, + 0.024862272664904594, + 0.007618390955030918, + 0.07979940623044968, + 0.10307180881500244, + 0.02306625433266163, + 0.010770892724394798, + 0.0071776551194489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.07002337276935577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3391542136669159, + 0.03904084861278534, + 0.029484733939170837, + 0.019007081165909767, + 0.03885156661272049, + 0.01027990784496069, + 0.05048706382513046, + 0.045134689658880234, + 0.28125816583633423, + 0.14730174839496613 + ], + "credit_map": [ + 0.13460977375507355, + 0.014667403884232044, + 0.1335659921169281, + 0.016976140439510345, + 0.0067459577694535255, + 0.0134799275547266, + 0.011257928796112537, + 0.007851089350879192, + 0.009883102029561996, + 0.004654292948544025, + 0.005497942678630352, + 0.004701097030192614, + 0.0017921606777235866, + 0.0031642429530620575, + 0.004250184632837772, + 0.00509782275184989, + 0.004983033053576946, + 0.005032507237046957, + 0.0037479179445654154, + 0.004241990856826305, + 0.0028919163160026073, + 0.00264286738820374, + 0.00251775118522346, + 0.0034267499577254057, + 0.003368949517607689, + 0.0026064070407301188, + 0.0026042312383651733, + 0.001718270592391491, + 0.0018092691898345947, + 0.0024740532971918583, + 0.0017467064317315817, + 0.0026791917625814676, + 0.004190683830529451, + 0.00921313650906086, + 0.0036123469471931458, + 0.0023454828187823296, + 0.00524721248075366, + 0.005366096273064613, + 0.002583863912150264, + 0.006292746867984533, + 0.0011154384119436145, + 0.0013223154237493873, + 0.0016827095532789826, + 0.0010702133877202868, + 0.0014671309618279338, + 0.0013781454181298614, + 0.0011482019908726215, + 0.0010957530466839671, + 0.0061743734404444695, + 0.005272580776363611, + 0.004539903718978167, + 0.004137805663049221, + 0.012230508960783482, + 0.007865331135690212, + 0.004274968523532152, + 0.005991591606289148, + 0.005221564322710037, + 0.0054310960695147514, + 0.0024559053126722574, + 0.0037048535887151957, + 0.008178615011274815, + 0.009463753551244736, + 0.006482310127466917, + 0.004196588881313801, + 0.017170671373605728, + 0.013217638246715069, + 0.010194460861384869, + 0.010268610902130604, + 0.013387990184128284, + 0.02061459794640541, + 0.09582538157701492, + 0.10057879239320755, + 0.00818631798028946, + 0.013417071662843227, + 0.027378682047128677, + 0.0070039816200733185, + 0.058663271367549896, + 0.015503334812819958, + 0.009267207235097885, + 0.007881879806518555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00409490754827857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, high-speed tracking shot, viewed from a slightly upward angle, closely follows a streamlined, futuristic spacecraft as it hurtles through the digital sky at breakneck speed. The vessel's hull, composed of a polished nanocomposite material resembling obsidian, is seamless and smooth, perfectly reflecting the light of the digital world below. Thin, cyan-blue energy lines flow along its aerodynamic contours, and its flanks are dotted with tiny, bright, pulsing navigation lights. The massive engines at the ship's rear belch blast hot blue plasma, creating a steady, dazzling trail of light. The force of this powerful thrust even slightly distorts the surrounding space. Beneath the ship, a vast, digital nightscape of the Guangdong-Hong Kong-Macao Greater Bay Area unfolds. This is not a physical city lightscape, but a three-dimensional holographic map composed of countless luminous blue and gold data streams. These data streams, like rushing rivers and highways, weave and pulsate through the virtual city skyline. As the spacecraft effortlessly zipped past at high speed, its trajectory created a fleeting ripple in these rivers of light. Digital particles, disturbed by its powerful aura, splattered to either side. Captured by a high-speed camera, the entire scene exhibited a breathtaking sense of dynamic beauty. The bright blue of the spacecraft's engines contrasted with the bluish-gold glow of the background data streams, creating a grand, futuristic atmosphere of high technology, extreme speed, and breathtaking speed.", + "seed": 18971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_11/index_6470_seed_18971_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0347679799888283, + "vlm_logprob": -0.0347679799888283, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0347679799888283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0498245470225811 + ], + "opposite_logprob_tokens": [ + -3.0498244762420654 + ], + "logprob_reward": -0.0498245470225811, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0498245470225811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12769773602485657, + 0.1052425354719162, + 0.05328413099050522, + 0.3378985524177551, + 0.06770063191652298, + 0.04583355039358139, + 0.0811891034245491, + 0.07586994022130966, + 0.04862974211573601, + 0.05665408447384834 + ], + "credit_map": [ + 0.026338504627346992, + 0.013101461343467236, + 0.008068561553955078, + 0.010872914455831051, + 0.025868043303489685, + 0.012461969628930092, + 0.01519853062927723, + 0.015787746757268906, + 0.016489170491695404, + 0.012514837086200714, + 0.006268765777349472, + 0.012665970250964165, + 0.006970027461647987, + 0.009839390404522419, + 0.028491148725152016, + 0.012003221549093723, + 0.0051393467001616955, + 0.006391643546521664, + 0.006236294284462929, + 0.009665559977293015, + 0.005364114884287119, + 0.006955801509320736, + 0.006950905080884695, + 0.006580462213605642, + 0.019049284979701042, + 0.014045454561710358, + 0.019202880561351776, + 0.08517375588417053, + 0.024429773911833763, + 0.08372633904218674, + 0.06742096692323685, + 0.024850133806467056, + 0.004964309278875589, + 0.006871484220027924, + 0.00741069670766592, + 0.006676078774034977, + 0.005766717251390219, + 0.006006378214806318, + 0.01612662896513939, + 0.013878334313631058, + 0.005453138612210751, + 0.005315143149346113, + 0.00617934949696064, + 0.006855217274278402, + 0.005206505302339792, + 0.00572594441473484, + 0.006562680471688509, + 0.004535572603344917, + 0.021281657740473747, + 0.012710441835224628, + 0.010103791020810604, + 0.0069246310740709305, + 0.0069877635687589645, + 0.007342976983636618, + 0.006550497375428677, + 0.009287345223128796, + 0.010698050260543823, + 0.010728942230343819, + 0.005338270682841539, + 0.0036795989144593477, + 0.013493595644831657, + 0.00460038660094142, + 0.005090139340609312, + 0.022240953519940376, + 0.006410412024706602, + 0.004899831488728523, + 0.005875265691429377, + 0.008431592024862766, + 0.006803017109632492, + 0.0037286286242306232, + 0.007946176454424858, + 0.004534818232059479, + 0.008223194628953934, + 0.011886867694556713, + 0.00873350165784359, + 0.007725741248577833, + 0.005838078446686268, + 0.0032459080684930086, + 0.006755517330020666, + 0.004245274234563112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0498245470225811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002837562235072255 + ], + "opposite_logprob_tokens": [ + -5.877837657928467 + ], + "logprob_reward": -0.002837562235072255, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002837562235072255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1462589055299759, + 0.0629599392414093, + 0.08497785776853561, + 0.2448849231004715, + 0.055934399366378784, + 0.055220238864421844, + 0.0984545424580574, + 0.08804462850093842, + 0.07020694017410278, + 0.09305767714977264 + ], + "credit_map": [ + 0.0205590408295393, + 0.016485348343849182, + 0.009658347815275192, + 0.014530801214277744, + 0.022360309958457947, + 0.01630818098783493, + 0.019038008525967598, + 0.027318879961967468, + 0.008154314942657948, + 0.008646450936794281, + 0.008318930864334106, + 0.008991393260657787, + 0.007531189359724522, + 0.007551595568656921, + 0.006980287376791239, + 0.00678577646613121, + 0.007027334067970514, + 0.010913302190601826, + 0.011726122349500656, + 0.012373357079923153, + 0.010044834576547146, + 0.008782311342656612, + 0.01218718383461237, + 0.011923416517674923, + 0.014655391685664654, + 0.011330819688737392, + 0.013107237406075, + 0.05980871245265007, + 0.02007611095905304, + 0.05973618105053902, + 0.05027245730161667, + 0.015897998586297035, + 0.005360665265470743, + 0.00649554468691349, + 0.007622533943504095, + 0.00823465920984745, + 0.006325311493128538, + 0.007498415187001228, + 0.007973159663379192, + 0.006424105726182461, + 0.005293691996484995, + 0.0063408613204956055, + 0.00904504768550396, + 0.006860337685793638, + 0.004461986944079399, + 0.007646611426025629, + 0.009023657068610191, + 0.006548046600073576, + 0.026625484228134155, + 0.016117319464683533, + 0.010551181621849537, + 0.008254400454461575, + 0.0071663702838122845, + 0.008634462021291256, + 0.007605499587953091, + 0.013499828055500984, + 0.00894123688340187, + 0.020245172083377838, + 0.006681223399937153, + 0.0056852009147405624, + 0.0233343243598938, + 0.006598785053938627, + 0.007500332780182362, + 0.009058353491127491, + 0.00770342955365777, + 0.004187629092484713, + 0.022080427035689354, + 0.015551615506410599, + 0.0045232330448925495, + 0.003369857557117939, + 0.008454736322164536, + 0.0043360101990401745, + 0.00792459025979042, + 0.018949387595057487, + 0.009112700819969177, + 0.012158588506281376, + 0.022238345816731453, + 0.006094626151025295, + 0.010467018000781536, + 0.006112419068813324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.002837562235072255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784241206943989 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784241206943989, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784241206943989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17020484805107117, + 0.11646231263875961, + 0.0676865428686142, + 0.10075239092111588, + 0.12651050090789795, + 0.06988883763551712, + 0.05488831549882889, + 0.05775846540927887, + 0.08063147962093353, + 0.1552163064479828 + ], + "credit_map": [ + 0.02968665398657322, + 0.015348534099757671, + 0.013106601312756538, + 0.017634868621826172, + 0.02905113436281681, + 0.015963589772582054, + 0.026413781568408012, + 0.022999677807092667, + 0.013308968394994736, + 0.014782968908548355, + 0.009901375509798527, + 0.01375827006995678, + 0.00730262091383338, + 0.01613652892410755, + 0.025249015539884567, + 0.016022568568587303, + 0.007461341097950935, + 0.008392440155148506, + 0.006458496209233999, + 0.008878752589225769, + 0.007776101119816303, + 0.010643073357641697, + 0.010051106102764606, + 0.008025229908525944, + 0.008432453498244286, + 0.009418880566954613, + 0.009876007214188576, + 0.01392924040555954, + 0.01154619175940752, + 0.01799427531659603, + 0.015190698206424713, + 0.014364645816385746, + 0.010616026818752289, + 0.008448397740721703, + 0.016405686736106873, + 0.011466195806860924, + 0.02048846147954464, + 0.012881416827440262, + 0.02225552126765251, + 0.02394879050552845, + 0.009428979828953743, + 0.007807271555066109, + 0.009944627992808819, + 0.009427954442799091, + 0.007445746101438999, + 0.010053013451397419, + 0.00931944977492094, + 0.006461792159825563, + 0.011100931093096733, + 0.005828184075653553, + 0.007232113741338253, + 0.005720383487641811, + 0.004571061581373215, + 0.007486744783818722, + 0.0057452102191746235, + 0.007203688845038414, + 0.0069326236844062805, + 0.006891126744449139, + 0.005589318461716175, + 0.004548569209873676, + 0.01228966936469078, + 0.0043177311308681965, + 0.005806377157568932, + 0.011383053846657276, + 0.006238853558897972, + 0.005311515647917986, + 0.022123483940958977, + 0.007885429076850414, + 0.005883791018277407, + 0.006039656698703766, + 0.02105700969696045, + 0.006091746501624584, + 0.013216778635978699, + 0.03568681329488754, + 0.014589721336960793, + 0.013313563540577888, + 0.04465106502175331, + 0.01112473476678133, + 0.012509283609688282, + 0.010124343447387218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009784241206943989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031916201114655 + ], + "opposite_logprob_tokens": [ + -2.3503191471099854 + ], + "logprob_reward": -0.10031916201114655, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031916201114655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.238469660282135, + 0.1581236720085144, + 0.06381433457136154, + 0.11946304142475128, + 0.05756744369864464, + 0.04960636794567108, + 0.1301792412996292, + 0.04839479923248291, + 0.03511819988489151, + 0.0992632657289505 + ], + "credit_map": [ + 0.047032181173563004, + 0.02096743695437908, + 0.014267845079302788, + 0.021526139229536057, + 0.04175280034542084, + 0.02332126721739769, + 0.031158220022916794, + 0.038443759083747864, + 0.025457555428147316, + 0.014496703632175922, + 0.007306593004614115, + 0.02333478257060051, + 0.010742822661995888, + 0.00898500345647335, + 0.04471202567219734, + 0.02308819070458412, + 0.006258014123886824, + 0.009669170714914799, + 0.006692472379654646, + 0.009362968616187572, + 0.007363014854490757, + 0.010013130493462086, + 0.007497815880924463, + 0.006957746110856533, + 0.007889773696660995, + 0.00942456815391779, + 0.009020054712891579, + 0.018101252615451813, + 0.01691489852964878, + 0.014769956469535828, + 0.029308725148439407, + 0.014033815823495388, + 0.006133894436061382, + 0.009147237055003643, + 0.007556178141385317, + 0.0066757784225046635, + 0.006557650398463011, + 0.006757688708603382, + 0.00741268415004015, + 0.007326329126954079, + 0.004717777483165264, + 0.00615198677405715, + 0.00771387480199337, + 0.006543521769344807, + 0.003953445702791214, + 0.007387533318251371, + 0.007467125076800585, + 0.0056710997596383095, + 0.038633547723293304, + 0.0195715744048357, + 0.011391276493668556, + 0.009713863953948021, + 0.009093644097447395, + 0.011856629513204098, + 0.013258268125355244, + 0.016660436987876892, + 0.0049048783257603645, + 0.009872953407466412, + 0.00858099665492773, + 0.0037806977052241564, + 0.010608959943056107, + 0.002433182904496789, + 0.004340676125138998, + 0.003872453700751066, + 0.0036799930967390537, + 0.003804117441177368, + 0.008940991014242172, + 0.006568270269781351, + 0.003417553612962365, + 0.0018279888899996877, + 0.004215514753013849, + 0.0026637704577296972, + 0.009996372275054455, + 0.01793750375509262, + 0.008677798323333263, + 0.011497443541884422, + 0.025925511494278908, + 0.0069115133956074715, + 0.011182315647602081, + 0.007134801242500544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.10031916201114655 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, futuristic spacecraft with a polished obsidian-like hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin cyan-blue energy lines visible flowing along the spacecraft's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the spacecraft's massive rear engines emit hot blue plasma creating a trail of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a three-dimensional holographic map of the Guangdong-Hong Kong-Macao Greater Bay Area composed of luminous blue and gold data streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a tight, high-speed tracking shot viewed from a slightly upward angle following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074387468397617 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074387468397617, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074387468397617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22125817835330963, + 0.084996297955513, + 0.07080788910388947, + 0.06718920171260834, + 0.10441157966852188, + 0.08301642537117004, + 0.05888889729976654, + 0.08711543679237366, + 0.17887328565120697, + 0.04344280809164047 + ], + "credit_map": [ + 0.029603293165564537, + 0.02723066322505474, + 0.014067555777728558, + 0.02302282489836216, + 0.03767397254705429, + 0.026069575920701027, + 0.027541346848011017, + 0.03604895621538162, + 0.008780073374509811, + 0.009478321298956871, + 0.007092698477208614, + 0.0091431038454175, + 0.010475777089595795, + 0.013562253676354885, + 0.015485596843063831, + 0.010978477075695992, + 0.009145687334239483, + 0.007915869355201721, + 0.005085345823317766, + 0.008398017846047878, + 0.009641056880354881, + 0.015803871676325798, + 0.005983832757920027, + 0.008834216743707657, + 0.00628996966406703, + 0.006314738187938929, + 0.007096486631780863, + 0.009032098576426506, + 0.005757881794124842, + 0.014066663570702076, + 0.007776082959026098, + 0.010855293832719326, + 0.005828447639942169, + 0.009994417428970337, + 0.009967905469238758, + 0.00799962505698204, + 0.009074847213923931, + 0.009692576713860035, + 0.02310134842991829, + 0.028752416372299194, + 0.010205981321632862, + 0.00784348789602518, + 0.016476793214678764, + 0.011209776625037193, + 0.007262124679982662, + 0.0075582717545330524, + 0.0143187390640378, + 0.008141256868839264, + 0.009050143882632256, + 0.006957783363759518, + 0.007838514633476734, + 0.006786254700273275, + 0.005733800586313009, + 0.0054147811606526375, + 0.009784426540136337, + 0.0073231919668614864, + 0.010383912362158298, + 0.019098443910479546, + 0.009205053560435772, + 0.005266113206744194, + 0.008040455169975758, + 0.005844923667609692, + 0.006015847437083721, + 0.02326068840920925, + 0.0239277221262455, + 0.012168681249022484, + 0.03892477974295616, + 0.04162309691309929, + 0.015451905317604542, + 0.013637679629027843, + 0.01983039267361164, + 0.013309040106832981, + 0.005490798037499189, + 0.006963147781789303, + 0.006841970607638359, + 0.0067046042531728745, + 0.005068386904895306, + 0.003661792492493987, + 0.00469919154420495, + 0.004012923687696457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.011074387468397617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_tensor([3067, 3067])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_tensor([3067, 3067])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..184ba79c52166ae12972f24b2d70ae35e25cc25d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_tensor([3067, 3067])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Through a canted Dutch angle lens, we gaze in close-up at a three-dimensional holographic weather map suspended within a dimly lit, futuristic command center. At the center of the frame, the energy volume representing the high-risk tornado zone appears as a viscous, luminous, scarlet non-Newtonian fluid, churning and tumbling violently within an invisible boundary. This luminous red mass moves in a chaotic yet organic manner, sometimes coalescing into a taut, swirling core, sometimes suddenly erupting outward in violent splashes. Countless fiery red droplets, as if captured by a high-speed camera, freeze in mid-air, their edges trailing fine strands of light. The scene is imbued with the dynamic, athletic style of cinematography, its sharp focus capturing every pulse of energy and every collision of fluids. The blurred blue grid lines and faint terrain contours in the background highlight the extreme instability of the red zone. The ominous glow emanating from the red fluid itself serves as the sole source of light, casting a tense red hue on the invisible surfaces of the surrounding control consoles, creating an atmosphere of intense urgency and imminent crisis.", + "seed": 23562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_3067_seed_23562_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23335887510329484, + "vlm_logprob": -0.23335887510329484, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.23335887510329484, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03737489879131317 + ], + "opposite_logprob_tokens": [ + -4.537374973297119 + ], + "logprob_reward": -0.03737489879131317, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03737489879131317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11404311656951904, + 0.12710583209991455, + 0.06558738648891449, + 0.05374260991811752, + 0.09760212153196335, + 0.09818866848945618, + 0.08284574747085571, + 0.1943918764591217, + 0.08497349172830582, + 0.08151914924383163 + ], + "credit_map": [ + 0.012767710722982883, + 0.013544976711273193, + 0.015069225803017616, + 0.01709485985338688, + 0.01392937172204256, + 0.012668006122112274, + 0.014292501844465733, + 0.014676460064947605, + 0.013591504655778408, + 0.02323281206190586, + 0.019197160378098488, + 0.013832411728799343, + 0.012759076431393623, + 0.015515993349254131, + 0.010570005513727665, + 0.018406875431537628, + 0.00670472951605916, + 0.007723939139395952, + 0.011284111998975277, + 0.011112095788121223, + 0.006895221769809723, + 0.006041268818080425, + 0.008952975273132324, + 0.006873048841953278, + 0.009033930487930775, + 0.007218796759843826, + 0.010018525645136833, + 0.0062149567529559135, + 0.0046652015298604965, + 0.005539681296795607, + 0.0060899583622813225, + 0.004961558151990175, + 0.008036463521420956, + 0.019127003848552704, + 0.007850339636206627, + 0.022378912195563316, + 0.01149036455899477, + 0.00781411211937666, + 0.012169948779046535, + 0.008734970353543758, + 0.008893923833966255, + 0.014554282650351524, + 0.013359198346734047, + 0.013253810815513134, + 0.01622605323791504, + 0.008599027059972286, + 0.014753776602447033, + 0.008548591285943985, + 0.006371662020683289, + 0.020238714292645454, + 0.010838508605957031, + 0.014676041901111603, + 0.009439471177756786, + 0.005644691176712513, + 0.010345163755118847, + 0.005291496403515339, + 0.017701521515846252, + 0.03301854059100151, + 0.0219286996871233, + 0.0227960254997015, + 0.028656648471951485, + 0.02870975434780121, + 0.019469942897558212, + 0.022110752761363983, + 0.008499953895807266, + 0.015003496780991554, + 0.00981635320931673, + 0.011618978343904018, + 0.012653673067688942, + 0.006452219095081091, + 0.014289281331002712, + 0.006639536935836077, + 0.01109680812805891, + 0.014327174052596092, + 0.01166493259370327, + 0.0070859030820429325, + 0.010387876071035862, + 0.009956777095794678, + 0.009880268014967442, + 0.007119408342987299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03737489879131317 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742931127548218 + ], + "opposite_logprob_tokens": [ + -0.4742931127548218 + ], + "logprob_reward": -0.9742931127548218, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742931127548218, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24161604046821594, + 0.09292072057723999, + 0.06754423677921295, + 0.05474726855754852, + 0.04463043436408043, + 0.1914689987897873, + 0.14592359960079193, + 0.049091119319200516, + 0.07109710574150085, + 0.04096047952771187 + ], + "credit_map": [ + 0.03104158490896225, + 0.030180972069501877, + 0.04093769192695618, + 0.028736552223563194, + 0.02517538145184517, + 0.031520843505859375, + 0.02387280762195587, + 0.030150189995765686, + 0.008797354064881802, + 0.02030615136027336, + 0.01188005693256855, + 0.00795110035687685, + 0.008786648511886597, + 0.013979443348944187, + 0.009407700970768929, + 0.011812262237071991, + 0.007045821752399206, + 0.0087512768805027, + 0.015085398219525814, + 0.008190474472939968, + 0.006878902204334736, + 0.008448831737041473, + 0.007311197929084301, + 0.005832336843013763, + 0.00608827592805028, + 0.008327887393534184, + 0.006996122654527426, + 0.007120940834283829, + 0.005719474516808987, + 0.008145605213940144, + 0.00565489474684, + 0.0066940663382411, + 0.002959122881293297, + 0.0142862843349576, + 0.0025166980922222137, + 0.01231383066624403, + 0.003084430703893304, + 0.002844969043508172, + 0.003506401553750038, + 0.0031187019776552916, + 0.020630288869142532, + 0.0315471850335598, + 0.01761668361723423, + 0.022414661943912506, + 0.025192508473992348, + 0.023842502385377884, + 0.025988133624196053, + 0.024237031117081642, + 0.010128426365554333, + 0.02505309134721756, + 0.015561139211058617, + 0.02983197383582592, + 0.011381890624761581, + 0.02549160271883011, + 0.012346827425062656, + 0.016128646209836006, + 0.004070914350450039, + 0.006366537418216467, + 0.008741105906665325, + 0.004643619991838932, + 0.007656235713511705, + 0.0058929044753313065, + 0.006124986335635185, + 0.005594814196228981, + 0.006671231240034103, + 0.015688687562942505, + 0.00974156241863966, + 0.009690250270068645, + 0.009184597060084343, + 0.00454130070284009, + 0.010227260179817677, + 0.005352210719138384, + 0.0034796593245118856, + 0.010181639343500137, + 0.011086872778832912, + 0.003480676095932722, + 0.004398736171424389, + 0.0019752869848161936, + 0.004348126705735922, + 0.00200948235578835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.9742931127548218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906880021095276 + ], + "opposite_logprob_tokens": [ + -2.46406888961792 + ], + "logprob_reward": -0.08906880021095276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906880021095276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10609082877635956, + 0.0730937272310257, + 0.05963858217000961, + 0.060517795383930206, + 0.05069722235202789, + 0.07803448289632797, + 0.08384153991937637, + 0.16880635917186737, + 0.2763523757457733, + 0.04292711615562439 + ], + "credit_map": [ + 0.011322678066790104, + 0.01406110730022192, + 0.018037259578704834, + 0.01217813789844513, + 0.01481801550835371, + 0.01082760188728571, + 0.012242423370480537, + 0.012603603303432465, + 0.009047119878232479, + 0.009169277735054493, + 0.012862799689173698, + 0.007963642477989197, + 0.008343320339918137, + 0.008223498240113258, + 0.009054956026375294, + 0.008429114706814289, + 0.005808885674923658, + 0.007478763349354267, + 0.009772134013473988, + 0.00976389367133379, + 0.006562432274222374, + 0.005906521808356047, + 0.00816075224429369, + 0.006185198202729225, + 0.007835878990590572, + 0.0077240243554115295, + 0.00614121975377202, + 0.008157940581440926, + 0.007871781475841999, + 0.008441301062703133, + 0.00770947290584445, + 0.0066361697390675545, + 0.0032371964771300554, + 0.010603743605315685, + 0.005273771472275257, + 0.009491759352385998, + 0.005107425153255463, + 0.004488865379244089, + 0.007491034921258688, + 0.0050034248270094395, + 0.008259980008006096, + 0.013343307189643383, + 0.0071114432066679, + 0.009712337516248226, + 0.009371488355100155, + 0.011130177415907383, + 0.008904671296477318, + 0.010201077908277512, + 0.0070258998312056065, + 0.015997517853975296, + 0.006460241507738829, + 0.012866305187344551, + 0.010429100133478642, + 0.013701451011002064, + 0.010004177689552307, + 0.0073568508960306644, + 0.017649881541728973, + 0.027309508994221687, + 0.018286097794771194, + 0.017830941826105118, + 0.028797928243875504, + 0.02228306420147419, + 0.017543591558933258, + 0.01910535804927349, + 0.02400672808289528, + 0.0590084046125412, + 0.03893602266907692, + 0.038874492049217224, + 0.04259767010807991, + 0.019881153479218483, + 0.03727860376238823, + 0.015769287943840027, + 0.004739365074783564, + 0.008899319916963577, + 0.009551351889967918, + 0.0024994283448904753, + 0.004627961199730635, + 0.004673071205615997, + 0.004391582682728767, + 0.0035450321156531572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.08906880021095276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084762401878834 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084762401878834, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084762401878834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08821438997983932, + 0.0582987405359745, + 0.05910048633813858, + 0.051834478974342346, + 0.069380983710289, + 0.1362803429365158, + 0.058609846979379654, + 0.07865981757640839, + 0.3458103537559509, + 0.05381055176258087 + ], + "credit_map": [ + 0.010402834042906761, + 0.014940598048269749, + 0.010055571794509888, + 0.011865573935210705, + 0.014326306991279125, + 0.007846450433135033, + 0.010882774367928505, + 0.007894282229244709, + 0.005541209131479263, + 0.007367988117039204, + 0.0074925110675394535, + 0.009290006943047047, + 0.006140114739537239, + 0.007539148908108473, + 0.008494025096297264, + 0.006433732341974974, + 0.007844637148082256, + 0.006895607803016901, + 0.011481688357889652, + 0.00747251883149147, + 0.006879935506731272, + 0.006390271242707968, + 0.0059583294205367565, + 0.006177498959004879, + 0.003998721018433571, + 0.009399189613759518, + 0.006681140046566725, + 0.010525570251047611, + 0.006025536917150021, + 0.003609849838539958, + 0.007016561925411224, + 0.0045779114589095116, + 0.0054937866516411304, + 0.012348689138889313, + 0.007940007373690605, + 0.01183317694813013, + 0.00755871320143342, + 0.005495068151503801, + 0.012624124996364117, + 0.006087419576942921, + 0.011187133379280567, + 0.020373353734612465, + 0.01553030963987112, + 0.01792934536933899, + 0.023972753435373306, + 0.01196852046996355, + 0.02131679281592369, + 0.014002133160829544, + 0.0036569698713719845, + 0.012450862675905228, + 0.004839278757572174, + 0.01756836473941803, + 0.004655264783650637, + 0.004990728106349707, + 0.006848035380244255, + 0.0036003452260047197, + 0.007553356699645519, + 0.010884162038564682, + 0.008230808191001415, + 0.008266769349575043, + 0.010141470469534397, + 0.01532676536589861, + 0.008709851652383804, + 0.009546639397740364, + 0.03020252101123333, + 0.07800239324569702, + 0.04404380917549133, + 0.050730813294649124, + 0.048630453646183014, + 0.024880267679691315, + 0.04984341561794281, + 0.019476672634482384, + 0.004980890080332756, + 0.007819890975952148, + 0.010101193562150002, + 0.00540170818567276, + 0.006014269310981035, + 0.006945942994207144, + 0.005957474932074547, + 0.00658918684348464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.011084762401878834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054972801357507706 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.054972801357507706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054972801357507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24772776663303375, + 0.0966079980134964, + 0.1120714321732521, + 0.054705072194337845, + 0.07716849446296692, + 0.10002432763576508, + 0.09665853530168533, + 0.04237745702266693, + 0.05097995325922966, + 0.12167893350124359 + ], + "credit_map": [ + 0.026151329278945923, + 0.05096025392413139, + 0.032070647925138474, + 0.03842928260564804, + 0.029975876212120056, + 0.020174292847514153, + 0.025506213307380676, + 0.024459870532155037, + 0.008449938148260117, + 0.020581116899847984, + 0.009893995709717274, + 0.014981621876358986, + 0.013956659473478794, + 0.008845281787216663, + 0.009800097905099392, + 0.010099289007484913, + 0.009889266453683376, + 0.014409438706934452, + 0.020124906674027443, + 0.020385725423693657, + 0.014032320119440556, + 0.010915347374975681, + 0.013836324214935303, + 0.008478100411593914, + 0.0049887909553945065, + 0.008989984169602394, + 0.01034598145633936, + 0.008261015638709068, + 0.0058375257067382336, + 0.004829589743167162, + 0.0060558076947927475, + 0.005396375432610512, + 0.005915642715990543, + 0.011277500540018082, + 0.007900512777268887, + 0.017522362992167473, + 0.01042940467596054, + 0.006096959114074707, + 0.011400173418223858, + 0.006625945679843426, + 0.01035058218985796, + 0.018150050193071365, + 0.008898760192096233, + 0.011299348436295986, + 0.012170313857495785, + 0.013372310437262058, + 0.0136281568557024, + 0.012154812924563885, + 0.005772438831627369, + 0.021064452826976776, + 0.009996442124247551, + 0.022127853706479073, + 0.009951509535312653, + 0.007224493194371462, + 0.014877472072839737, + 0.005643873475492001, + 0.005437426269054413, + 0.00513260904699564, + 0.006470371037721634, + 0.005075990222394466, + 0.005198470316827297, + 0.004696143791079521, + 0.006001776549965143, + 0.0043646665289998055, + 0.004906254354864359, + 0.009309347718954086, + 0.008076966740190983, + 0.00628119008615613, + 0.007050032261759043, + 0.004470836836844683, + 0.006812308914959431, + 0.004073013551533222, + 0.011018424294888973, + 0.022662431001663208, + 0.034670088440179825, + 0.00870467908680439, + 0.012393870390951633, + 0.009741991758346558, + 0.0113332225009799, + 0.011154226027429104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.054972801357507706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a canted Dutch angle lens, we gaze in close-up at a three-dimensional holographic weather map suspended within a dimly lit, futuristic command center. At the center of the frame, the energy volume representing the high-risk tornado zone appears as a viscous, luminous, scarlet non-Newtonian fluid, churning and tumbling violently within an invisible boundary. This luminous red mass moves in a chaotic yet organic manner, sometimes coalescing into a taut, swirling core, sometimes suddenly erupting outward in violent splashes. Countless fiery red droplets, as if captured by a high-speed camera, freeze in mid-air, their edges trailing fine strands of light. The scene is imbued with the dynamic, athletic style of cinematography, its sharp focus capturing every pulse of energy and every collision of fluids. The blurred blue grid lines and faint terrain contours in the background highlight the extreme instability of the red zone. The ominous glow emanating from the red fluid itself serves as the sole source of light, casting a tense red hue on the invisible surfaces of the surrounding control consoles, creating an atmosphere of intense urgency and imminent crisis.", + "seed": 23563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_3067_seed_23563_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30619050869718195, + "vlm_logprob": -0.30619050869718195, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.30619050869718195, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041287507861852646 + ], + "opposite_logprob_tokens": [ + -4.916287422180176 + ], + "logprob_reward": -0.041287507861852646, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.041287507861852646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07817117869853973, + 0.30747830867767334, + 0.04574846848845482, + 0.03451291099190712, + 0.02562384493649006, + 0.021110964938998222, + 0.09646239131689072, + 0.17347784340381622, + 0.016302285715937614, + 0.2011118233203888 + ], + "credit_map": [ + 0.007854875177145004, + 0.005142900627106428, + 0.008276807144284248, + 0.01908191666007042, + 0.007419776637107134, + 0.00930565781891346, + 0.011356484144926071, + 0.00973274651914835, + 0.01753433793783188, + 0.015199686400592327, + 0.0265058446675539, + 0.0760292112827301, + 0.017855962738394737, + 0.04162203148007393, + 0.08225926011800766, + 0.030471965670585632, + 0.004356907214969397, + 0.005956009496003389, + 0.004187425132840872, + 0.010579799301922321, + 0.0049169897101819515, + 0.004022056236863136, + 0.007451560348272324, + 0.004277714993804693, + 0.0033678922336548567, + 0.004006678704172373, + 0.00514267897233367, + 0.005719715263694525, + 0.003596386406570673, + 0.0039118207059800625, + 0.004410646855831146, + 0.004357088357210159, + 0.0018243716331198812, + 0.0026577787939459085, + 0.003953919280320406, + 0.005052139982581139, + 0.002322267508134246, + 0.002335629891604185, + 0.004759360570460558, + 0.00271837436594069, + 0.0016135290497913957, + 0.00221991422586143, + 0.002984353806823492, + 0.0030284272506833076, + 0.00205979123711586, + 0.001983225578442216, + 0.003944687079638243, + 0.0032770324032753706, + 0.007176544982939959, + 0.004758013878017664, + 0.0071773589588701725, + 0.01331893727183342, + 0.005343068856745958, + 0.02462981455028057, + 0.025911036878824234, + 0.008147606626152992, + 0.009429734200239182, + 0.005447073839604855, + 0.06847511231899261, + 0.021771356463432312, + 0.0055722580291330814, + 0.013442421332001686, + 0.040477290749549866, + 0.008862585760653019, + 0.0010860534384846687, + 0.0023556130472570658, + 0.0022829235531389713, + 0.004303453024476767, + 0.001248766202479601, + 0.0013202889822423458, + 0.002298800041899085, + 0.0014063852140679955, + 0.016876477748155594, + 0.019888216629624367, + 0.027503667399287224, + 0.040797725319862366, + 0.015795081853866577, + 0.0248864758759737, + 0.02932576835155487, + 0.026038387790322304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.041287507861852646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.223615050315857 + ], + "opposite_logprob_tokens": [ + -0.34861502051353455 + ], + "logprob_reward": -1.223615050315857, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.223615050315857, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14081689715385437, + 0.16425420343875885, + 0.07269179821014404, + 0.06987085193395615, + 0.03229900449514389, + 0.10265769064426422, + 0.12522433698177338, + 0.15835930407047272, + 0.014016679488122463, + 0.11980928480625153 + ], + "credit_map": [ + 0.014810334891080856, + 0.007930134423077106, + 0.017454715445637703, + 0.037227995693683624, + 0.008932583965361118, + 0.013921414501965046, + 0.023329410701990128, + 0.01721029542386532, + 0.010833244770765305, + 0.006284436676651239, + 0.013811366632580757, + 0.03197940066456795, + 0.008246725425124168, + 0.020177865400910378, + 0.052109118551015854, + 0.020812034606933594, + 0.007630007341504097, + 0.011306636035442352, + 0.0103446114808321, + 0.00995669886469841, + 0.00872393511235714, + 0.006896620150655508, + 0.01029537059366703, + 0.007537917699664831, + 0.006886128336191177, + 0.007917712442576885, + 0.007408199831843376, + 0.013453972525894642, + 0.006869209464639425, + 0.0062584239058196545, + 0.013466846197843552, + 0.00761035131290555, + 0.0021953927353024483, + 0.002796166343614459, + 0.004036861006170511, + 0.0073896367102861404, + 0.003283047815784812, + 0.002454916015267372, + 0.0065774330869317055, + 0.0035655470564961433, + 0.009321415796875954, + 0.00661824643611908, + 0.014934880658984184, + 0.013023707084357738, + 0.007243738509714603, + 0.01175967138260603, + 0.014202877879142761, + 0.025553135201334953, + 0.008577780798077583, + 0.0049635274335742, + 0.008592123165726662, + 0.01814103126525879, + 0.006256348919123411, + 0.02918829210102558, + 0.04024572670459747, + 0.009259494952857494, + 0.009197121486067772, + 0.006183056626468897, + 0.06152091175317764, + 0.022476546466350555, + 0.005013021640479565, + 0.014782044105231762, + 0.02993057109415531, + 0.009256021119654179, + 0.0014047193108126521, + 0.0016225812723860145, + 0.0020463878754526377, + 0.0017632396193221211, + 0.0012461686274036765, + 0.002171766711398959, + 0.001775017473846674, + 0.0019867976661771536, + 0.00789519865065813, + 0.008924254216253757, + 0.011876069009304047, + 0.02912561595439911, + 0.011082026176154613, + 0.010657260194420815, + 0.026863103732466698, + 0.013385741040110588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -1.223615050315857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559872269630432 + ], + "opposite_logprob_tokens": [ + -1.600598692893982 + ], + "logprob_reward": -0.22559872269630432, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559872269630432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17315158247947693, + 0.08346402645111084, + 0.13045160472393036, + 0.1020415648818016, + 0.032279420644044876, + 0.15638935565948486, + 0.06012061610817909, + 0.106888048350811, + 0.036824800074100494, + 0.11838898062705994 + ], + "credit_map": [ + 0.017700064927339554, + 0.021285047754645348, + 0.018860215321183205, + 0.025334827601909637, + 0.031011082231998444, + 0.017200399190187454, + 0.0209999680519104, + 0.020759981125593185, + 0.0074054584838449955, + 0.003307587932795286, + 0.0048248982056975365, + 0.01990526169538498, + 0.0037816951517015696, + 0.012102431617677212, + 0.02522215247154236, + 0.0069145457819104195, + 0.01413039118051529, + 0.02198207564651966, + 0.014202355407178402, + 0.018469490110874176, + 0.017794158309698105, + 0.01325985137373209, + 0.01687144674360752, + 0.013741838745772839, + 0.01010005921125412, + 0.009483185596764088, + 0.011328312568366528, + 0.024115756154060364, + 0.009215586818754673, + 0.009569685906171799, + 0.01644327864050865, + 0.01178570557385683, + 0.0027404665015637875, + 0.002338846679776907, + 0.004516266752034426, + 0.008040115237236023, + 0.002800813177600503, + 0.0025081883650273085, + 0.0052277580834925175, + 0.004106967709958553, + 0.01316455565392971, + 0.008049342781305313, + 0.02223300188779831, + 0.018836336210370064, + 0.010150203481316566, + 0.015523753128945827, + 0.021294662728905678, + 0.04713749885559082, + 0.003820565761998296, + 0.003184448927640915, + 0.008359757252037525, + 0.009781225584447384, + 0.0032551174517720938, + 0.007716348394751549, + 0.017234649509191513, + 0.00676850276067853, + 0.006150911096483469, + 0.004100409336388111, + 0.01915588416159153, + 0.019986433908343315, + 0.003967583179473877, + 0.02168668434023857, + 0.024344073608517647, + 0.007496072445064783, + 0.0031847006175667048, + 0.0031477841548621655, + 0.0048319329507648945, + 0.005884765647351742, + 0.003260991768911481, + 0.005578603129833937, + 0.006054275203496218, + 0.00488174706697464, + 0.00944562442600727, + 0.008963733911514282, + 0.01607785001397133, + 0.02478032000362873, + 0.01111859641969204, + 0.014232250861823559, + 0.01916344091296196, + 0.014607161283493042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.22559872269630432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060079045593738556, + 0.1264704167842865, + 0.0642748773097992, + 0.031314417719841, + 0.06081223487854004, + 0.08549079298973083, + 0.03155871853232384, + 0.47110188007354736, + 0.021151868626475334, + 0.0477457158267498 + ], + "credit_map": [ + 0.006876319646835327, + 0.00533917685970664, + 0.010384883731603622, + 0.00930470135062933, + 0.005725247785449028, + 0.007762230932712555, + 0.005936351604759693, + 0.008750136941671371, + 0.006315677426755428, + 0.004342225845903158, + 0.006610898766666651, + 0.03367100656032562, + 0.005669737234711647, + 0.021059192717075348, + 0.017737600952386856, + 0.03106408752501011, + 0.004607536364346743, + 0.005282195750623941, + 0.008862193673849106, + 0.011376638896763325, + 0.009769472293555737, + 0.0050694914534688, + 0.012664873152971268, + 0.006642479915171862, + 0.002682196209207177, + 0.0038025458343327045, + 0.004087948240339756, + 0.006446898449212313, + 0.003148458432406187, + 0.0026866148691624403, + 0.005094648338854313, + 0.0033651133999228477, + 0.004250211641192436, + 0.0077104936353862286, + 0.009267996065318584, + 0.010702939704060555, + 0.0072874207980930805, + 0.005631035193800926, + 0.007452094927430153, + 0.0085100457072258, + 0.007029934320598841, + 0.005056709051132202, + 0.010580669157207012, + 0.011512194760143757, + 0.006772639229893684, + 0.007423565723001957, + 0.019023295491933823, + 0.01809179037809372, + 0.0024578545708209276, + 0.0017738478491082788, + 0.0031777366530150175, + 0.004337162710726261, + 0.0018468573689460754, + 0.008283150382339954, + 0.0067463223822414875, + 0.0029357888270169497, + 0.027803722769021988, + 0.011892123147845268, + 0.18868526816368103, + 0.06629225611686707, + 0.01353021152317524, + 0.05642930045723915, + 0.08173538744449615, + 0.02473367378115654, + 0.0018737976206466556, + 0.002116622868925333, + 0.003366647055372596, + 0.0029998202808201313, + 0.002173199551180005, + 0.00253482349216938, + 0.002958692377433181, + 0.00312826712615788, + 0.002953484421595931, + 0.004525338765233755, + 0.006139946635812521, + 0.015093338675796986, + 0.0038921909872442484, + 0.0030926293693482876, + 0.006706071086227894, + 0.005342717748135328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.006741635967046022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370962664484978 + ], + "opposite_logprob_tokens": [ + -3.4087095260620117 + ], + "logprob_reward": -0.03370962664484978, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370962664484978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14462162554264069, + 0.16246965527534485, + 0.02054043859243393, + 0.03573138266801834, + 0.07299122214317322, + 0.10712990909814835, + 0.022553278133273125, + 0.3573734760284424, + 0.023167280480265617, + 0.05342170596122742 + ], + "credit_map": [ + 0.01167912594974041, + 0.013572024181485176, + 0.020940344780683517, + 0.03306625410914421, + 0.015910230576992035, + 0.014095600694417953, + 0.017810633406043053, + 0.017547419294714928, + 0.010396448895335197, + 0.009088439866900444, + 0.010643093846738338, + 0.025251397863030434, + 0.011824177578091621, + 0.020016932860016823, + 0.025990745052695274, + 0.049258437007665634, + 0.0021998537704348564, + 0.0018947042990475893, + 0.003112811129540205, + 0.0030582973267883062, + 0.0029213912785053253, + 0.00247845146805048, + 0.0025986023247241974, + 0.002276327693834901, + 0.0034350459463894367, + 0.003306679893285036, + 0.003908777609467506, + 0.009356437250971794, + 0.002840358531102538, + 0.004095058422535658, + 0.005335867404937744, + 0.0034531601704657078, + 0.005458117928355932, + 0.009605856612324715, + 0.010296686552464962, + 0.012954585254192352, + 0.008256372064352036, + 0.007525240071117878, + 0.008544739335775375, + 0.010349631309509277, + 0.006822075694799423, + 0.006700410973280668, + 0.015588155947625637, + 0.014157403260469437, + 0.009732692502439022, + 0.008577817119657993, + 0.021548820659518242, + 0.02400253713130951, + 0.0015934989787638187, + 0.001759727718308568, + 0.0026670468505471945, + 0.003987169358879328, + 0.00138366618193686, + 0.003956317435950041, + 0.005281561054289341, + 0.0019242927664890885, + 0.02107131853699684, + 0.013632099144160748, + 0.13900750875473022, + 0.047160714864730835, + 0.011604726314544678, + 0.03687014430761337, + 0.06848376989364624, + 0.01954321190714836, + 0.0017938363598659635, + 0.00208981242030859, + 0.003269550856202841, + 0.003048212267458439, + 0.0019807282369583845, + 0.0034665262792259455, + 0.004823877941817045, + 0.0026947378646582365, + 0.003930277191102505, + 0.005068907048553228, + 0.008097359910607338, + 0.008906934410333633, + 0.004775427281856537, + 0.008226295001804829, + 0.006896537728607655, + 0.007519968785345554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03370962664484978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a canted Dutch angle lens, we gaze in close-up at a three-dimensional holographic weather map suspended within a dimly lit, futuristic command center. At the center of the frame, the energy volume representing the high-risk tornado zone appears as a viscous, luminous, scarlet non-Newtonian fluid, churning and tumbling violently within an invisible boundary. This luminous red mass moves in a chaotic yet organic manner, sometimes coalescing into a taut, swirling core, sometimes suddenly erupting outward in violent splashes. Countless fiery red droplets, as if captured by a high-speed camera, freeze in mid-air, their edges trailing fine strands of light. The scene is imbued with the dynamic, athletic style of cinematography, its sharp focus capturing every pulse of energy and every collision of fluids. The blurred blue grid lines and faint terrain contours in the background highlight the extreme instability of the red zone. The ominous glow emanating from the red fluid itself serves as the sole source of light, casting a tense red hue on the invisible surfaces of the surrounding control consoles, creating an atmosphere of intense urgency and imminent crisis.", + "seed": 23564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_3067_seed_23564_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.31645894795656204, + "vlm_logprob": -0.31645894795656204, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.31645894795656204, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03150901198387146 + ], + "opposite_logprob_tokens": [ + -6.406508922576904 + ], + "logprob_reward": -0.03150901198387146, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.03150901198387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17425566911697388, + 0.03456627577543259, + 0.06367504596710205, + 0.06762265413999557, + 0.24731048941612244, + 0.05490246042609215, + 0.10251720249652863, + 0.10263650119304657, + 0.05651106312870979, + 0.09600261598825455 + ], + "credit_map": [ + 0.02032468281686306, + 0.01926601119339466, + 0.027071814984083176, + 0.023119395598769188, + 0.016620751470327377, + 0.022335074841976166, + 0.020015720278024673, + 0.02550223469734192, + 0.0024488784838467836, + 0.004570494405925274, + 0.004497241228818893, + 0.00404087919741869, + 0.004264135379344225, + 0.006175365764647722, + 0.0036365871783345938, + 0.00493270019069314, + 0.005835023242980242, + 0.009474663063883781, + 0.00643924530595541, + 0.013445837423205376, + 0.008118894882500172, + 0.008813388645648956, + 0.005475874990224838, + 0.006072119809687138, + 0.005775531753897667, + 0.00919532310217619, + 0.009386489167809486, + 0.017663775011897087, + 0.008005051873624325, + 0.006932490039616823, + 0.004786502104252577, + 0.005877496208995581, + 0.021232323721051216, + 0.03156116604804993, + 0.03542950749397278, + 0.04204881191253662, + 0.024248581379652023, + 0.030510589480400085, + 0.015988789498806, + 0.04629075899720192, + 0.003971578553318977, + 0.008954091928899288, + 0.009839215315878391, + 0.00707671232521534, + 0.008876531384885311, + 0.005073469132184982, + 0.005020718090236187, + 0.006090151611715555, + 0.005464533343911171, + 0.026687124744057655, + 0.007741803303360939, + 0.013240781612694263, + 0.00787541177123785, + 0.017146874219179153, + 0.005818441975861788, + 0.01854224130511284, + 0.005378936417400837, + 0.007145564071834087, + 0.008779800496995449, + 0.029470980167388916, + 0.012041955254971981, + 0.018296822905540466, + 0.00812339223921299, + 0.01339905895292759, + 0.00470838975161314, + 0.009895986877381802, + 0.007320656906813383, + 0.006865945644676685, + 0.008761744946241379, + 0.004892297554761171, + 0.0062566776759922504, + 0.0078093670308589935, + 0.0057370769791305065, + 0.023026220500469208, + 0.010575216263532639, + 0.01700880378484726, + 0.01235343050211668, + 0.008670235984027386, + 0.009208869189023972, + 0.00942277256399393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03150901198387146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538768768310547 + ], + "opposite_logprob_tokens": [ + -0.4288769066333771 + ], + "logprob_reward": -1.0538768768310547, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538768768310547, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1838553547859192, + 0.0836549922823906, + 0.02834896184504032, + 0.03919769823551178, + 0.3696613311767578, + 0.04557603970170021, + 0.04822475090622902, + 0.14617860317230225, + 0.023448925465345383, + 0.03185326233506203 + ], + "credit_map": [ + 0.027987368404865265, + 0.014653476886451244, + 0.022271370515227318, + 0.01466649305075407, + 0.014104235917329788, + 0.019520319998264313, + 0.0120870815590024, + 0.05856500193476677, + 0.01143404096364975, + 0.010293385945260525, + 0.011038254015147686, + 0.009890539571642876, + 0.008664405904710293, + 0.011487684212625027, + 0.00927484780550003, + 0.011571832932531834, + 0.001957500819116831, + 0.0037039569579064846, + 0.0029472731985151768, + 0.005297569092363119, + 0.003415418555960059, + 0.005322282202541828, + 0.0024108700454235077, + 0.003294091671705246, + 0.0030680797062814236, + 0.004889240488409996, + 0.003695956664159894, + 0.008390875533223152, + 0.004030773416161537, + 0.006135466508567333, + 0.002434855792671442, + 0.006552450824528933, + 0.032116297632455826, + 0.04503433406352997, + 0.04710591956973076, + 0.06006408482789993, + 0.03517390787601471, + 0.04530435800552368, + 0.024107079952955246, + 0.08075538277626038, + 0.0031465443316847086, + 0.0059160348027944565, + 0.008859485387802124, + 0.005240029655396938, + 0.005551599431782961, + 0.00624579843133688, + 0.0037559978663921356, + 0.006860552355647087, + 0.0036448808386921883, + 0.006850087083876133, + 0.006813598796725273, + 0.008791918866336346, + 0.004168834537267685, + 0.008190130814909935, + 0.0036682707723230124, + 0.006097032222896814, + 0.005791171453893185, + 0.01060070563107729, + 0.009545103646814823, + 0.05067305266857147, + 0.01655750535428524, + 0.02635679766535759, + 0.012337482534348965, + 0.014316783286631107, + 0.0016136872582137585, + 0.0047064111568033695, + 0.002639210317283869, + 0.002852803561836481, + 0.00397759024053812, + 0.0029437378980219364, + 0.0020165315363556147, + 0.0026989553589373827, + 0.0036888429895043373, + 0.0045007928274571896, + 0.004165864549577236, + 0.003929832950234413, + 0.0038994145579636097, + 0.004134927876293659, + 0.0038038177881389856, + 0.0037297673989087343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -1.0538768768310547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288890063762665 + ], + "opposite_logprob_tokens": [ + -1.0538890361785889 + ], + "logprob_reward": -0.4288890063762665, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288890063762665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13586977124214172, + 0.04042172431945801, + 0.013860262930393219, + 0.022112004458904266, + 0.47731685638427734, + 0.10685409605503082, + 0.0092096496373415, + 0.0947108119726181, + 0.017112715169787407, + 0.08253207057714462 + ], + "credit_map": [ + 0.009780682623386383, + 0.008488805964589119, + 0.025796696543693542, + 0.010955298319458961, + 0.008616040460765362, + 0.024690834805369377, + 0.009533694013953209, + 0.03800772503018379, + 0.0035519336815923452, + 0.004491239786148071, + 0.006282174494117498, + 0.0068655977956950665, + 0.002972559304907918, + 0.00495646009221673, + 0.002485166536644101, + 0.008816594257950783, + 0.0016681523993611336, + 0.0020052201580256224, + 0.0013679431285709143, + 0.0024911866057664156, + 0.0018869952764362097, + 0.001373932813294232, + 0.0012190743582323194, + 0.0018477588891983032, + 0.0023243757896125317, + 0.0025889899116009474, + 0.003535148687660694, + 0.003945063799619675, + 0.00231837946921587, + 0.002876672660931945, + 0.0014612034428864717, + 0.003062172792851925, + 0.04049067571759224, + 0.056057773530483246, + 0.059711091220378876, + 0.08026096224784851, + 0.044676221907138824, + 0.056768544018268585, + 0.031353309750556946, + 0.1079983338713646, + 0.00931131187826395, + 0.012785758823156357, + 0.01707720197737217, + 0.013056956231594086, + 0.009961524046957493, + 0.016286028549075127, + 0.008674051612615585, + 0.01970127783715725, + 0.0006461555603891611, + 0.001552910660393536, + 0.001176406629383564, + 0.0018642869545146823, + 0.0010121858213096857, + 0.0012347480515018106, + 0.0008020598324947059, + 0.0009208968840539455, + 0.0035041777882725, + 0.007257508579641581, + 0.005307632032781839, + 0.034483615309000015, + 0.011001038365066051, + 0.016449740156531334, + 0.007899201475083828, + 0.008807907812297344, + 0.0014864910626783967, + 0.0022408845834434032, + 0.0019020579056814313, + 0.0021340155508369207, + 0.0027445252053439617, + 0.0022847431246191263, + 0.001994987018406391, + 0.0023250121157616377, + 0.008681912906467915, + 0.01573787070810795, + 0.00900988932698965, + 0.01297274511307478, + 0.010544821619987488, + 0.008874425664544106, + 0.007467400282621384, + 0.009243004955351353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.4288890063762665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954977124929428 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954977124929428, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954977124929428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470339447259903, + 0.056387078016996384, + 0.03310031816363335, + 0.09884154796600342, + 0.20948615670204163, + 0.08702459186315536, + 0.05586874857544899, + 0.13494476675987244, + 0.08259891718626022, + 0.09471387416124344 + ], + "credit_map": [ + 0.01010898593813181, + 0.016120899468660355, + 0.026322532445192337, + 0.017009301111102104, + 0.011808469891548157, + 0.01885507069528103, + 0.013003102503716946, + 0.033805590122938156, + 0.0052411481738090515, + 0.0066039469093084335, + 0.007065699435770512, + 0.00689186155796051, + 0.007867281325161457, + 0.008358344435691833, + 0.007203630171716213, + 0.0071551683358848095, + 0.0031232615001499653, + 0.00706119230017066, + 0.00324311968870461, + 0.005532483570277691, + 0.004003599751740694, + 0.0038230547215789557, + 0.0027743626851588488, + 0.003539250697940588, + 0.007339616771787405, + 0.013797145336866379, + 0.013610689900815487, + 0.018241576850414276, + 0.011131194420158863, + 0.013424022123217583, + 0.007384364493191242, + 0.013912945054471493, + 0.017824195325374603, + 0.025847680866718292, + 0.025318188592791557, + 0.030502425506711006, + 0.018931202590465546, + 0.025783861055970192, + 0.013510788790881634, + 0.051767829805612564, + 0.006582643371075392, + 0.009973966516554356, + 0.01454678364098072, + 0.010731551796197891, + 0.01139802299439907, + 0.011161192320287228, + 0.009357786737382412, + 0.013272656127810478, + 0.002493538660928607, + 0.018401633948087692, + 0.003681307192891836, + 0.009354811161756516, + 0.00397223886102438, + 0.010055656544864178, + 0.002997307339683175, + 0.004912259988486767, + 0.005995891522616148, + 0.010564012452960014, + 0.011063648387789726, + 0.0432782843708992, + 0.014706622809171677, + 0.0219915509223938, + 0.011150828562676907, + 0.016193920746445656, + 0.00710663478821516, + 0.012741923332214355, + 0.01114723365753889, + 0.012601668015122414, + 0.010944336652755737, + 0.009428011253476143, + 0.007410096935927868, + 0.011219018138945103, + 0.00876539945602417, + 0.022547278553247452, + 0.010449336841702461, + 0.012432942166924477, + 0.012920617125928402, + 0.008834799751639366, + 0.007228786591440439, + 0.011534718796610832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.005954977124929428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062064867466688156 + ], + "opposite_logprob_tokens": [ + -2.8120648860931396 + ], + "logprob_reward": -0.062064867466688156, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062064867466688156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09409315139055252, + 0.03812642768025398, + 0.025714796036481857, + 0.02110707014799118, + 0.22639557719230652, + 0.08502385020256042, + 0.10804150998592377, + 0.1904240995645523, + 0.04286065325140953, + 0.1682128608226776 + ], + "credit_map": [ + 0.006037162151187658, + 0.006319263484328985, + 0.015529236756265163, + 0.007915821857750416, + 0.0064541045576334, + 0.017268802970647812, + 0.005524973850697279, + 0.029043791815638542, + 0.004105699248611927, + 0.003896682523190975, + 0.0050660897977650166, + 0.00588810071349144, + 0.00410712044686079, + 0.005867787171155214, + 0.0039490461349487305, + 0.005245900247246027, + 0.0035108113661408424, + 0.0031038890592753887, + 0.0028880045283585787, + 0.0046552312560379505, + 0.0032222489826381207, + 0.002753137843683362, + 0.0021105946507304907, + 0.00347087811678648, + 0.002063133753836155, + 0.0031430392991751432, + 0.002983513055369258, + 0.00318371644243598, + 0.002438917523249984, + 0.002801933092996478, + 0.002386613516137004, + 0.0021062043961137533, + 0.020449601113796234, + 0.029169436544179916, + 0.03241686522960663, + 0.02927166037261486, + 0.0209528636187315, + 0.02925199270248413, + 0.015506704337894917, + 0.0493764728307724, + 0.006001024041324854, + 0.010332208126783371, + 0.01482941024005413, + 0.00806794036179781, + 0.014389277435839176, + 0.010142392478883266, + 0.010669147595763206, + 0.010592452250421047, + 0.005007544998079538, + 0.0381183847784996, + 0.0037069525569677353, + 0.01288510300219059, + 0.008147375658154488, + 0.024510443210601807, + 0.006128185894340277, + 0.00953751988708973, + 0.0070077041164040565, + 0.015666961669921875, + 0.01208848413079977, + 0.06617560982704163, + 0.021723268553614616, + 0.03353156894445419, + 0.015763482078909874, + 0.018467003479599953, + 0.00281312083825469, + 0.006673546973615885, + 0.005915224086493254, + 0.0063926405273377895, + 0.007304856553673744, + 0.004637186881154776, + 0.0039865425787866116, + 0.005137534346431494, + 0.021292047575116158, + 0.025666413828730583, + 0.02127472311258316, + 0.02455112524330616, + 0.021045280620455742, + 0.01689722388982773, + 0.015078574419021606, + 0.022407464683055878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.062064867466688156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a canted Dutch angle lens, we gaze in close-up at a three-dimensional holographic weather map suspended within a dimly lit, futuristic command center. At the center of the frame, the energy volume representing the high-risk tornado zone appears as a viscous, luminous, scarlet non-Newtonian fluid, churning and tumbling violently within an invisible boundary. This luminous red mass moves in a chaotic yet organic manner, sometimes coalescing into a taut, swirling core, sometimes suddenly erupting outward in violent splashes. Countless fiery red droplets, as if captured by a high-speed camera, freeze in mid-air, their edges trailing fine strands of light. The scene is imbued with the dynamic, athletic style of cinematography, its sharp focus capturing every pulse of energy and every collision of fluids. The blurred blue grid lines and faint terrain contours in the background highlight the extreme instability of the red zone. The ominous glow emanating from the red fluid itself serves as the sole source of light, casting a tense red hue on the invisible surfaces of the surrounding control consoles, creating an atmosphere of intense urgency and imminent crisis.", + "seed": 23565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_12/index_3067_seed_23565_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27561760433018206, + "vlm_logprob": -0.27561760433018206, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.27561760433018206, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05033217743039131 + ], + "opposite_logprob_tokens": [ + -3.9253320693969727 + ], + "logprob_reward": -0.05033217743039131, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05033217743039131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10858791321516037, + 0.06978056579828262, + 0.17375338077545166, + 0.05814724043011665, + 0.03131946548819542, + 0.03185758367180824, + 0.14344371855258942, + 0.16291634738445282, + 0.17123150825500488, + 0.04896223917603493 + ], + "credit_map": [ + 0.015905750915408134, + 0.012691881507635117, + 0.011770665645599365, + 0.014447096735239029, + 0.023920033127069473, + 0.007728842552751303, + 0.012942626141011715, + 0.009181019850075245, + 0.013582701794803143, + 0.008765850216150284, + 0.00446621049195528, + 0.003882828401401639, + 0.0243842750787735, + 0.004923072177916765, + 0.004446874838322401, + 0.005328753963112831, + 0.03288961574435234, + 0.016656741499900818, + 0.009459223598241806, + 0.014652974903583527, + 0.06600048393011093, + 0.011809008195996284, + 0.009970544837415218, + 0.012314791791141033, + 0.005665016360580921, + 0.005365652497857809, + 0.004286672919988632, + 0.01723855920135975, + 0.008278866298496723, + 0.005087914876639843, + 0.005336848087608814, + 0.0068877157755196095, + 0.0033678896725177765, + 0.00499751977622509, + 0.003921426832675934, + 0.005470322910696268, + 0.004480395931750536, + 0.0024038415867835283, + 0.0033484352752566338, + 0.0033296362962573767, + 0.0028140353970229626, + 0.00407452555373311, + 0.0030792937614023685, + 0.0038649996276944876, + 0.007275174837559462, + 0.0036342458333820105, + 0.004860653076320887, + 0.0022546558175235987, + 0.012965192086994648, + 0.01906070113182068, + 0.007888752035796642, + 0.05251637473702431, + 0.016537075862288475, + 0.008266556076705456, + 0.008922135457396507, + 0.017286919057369232, + 0.013760188594460487, + 0.013648184016346931, + 0.009687375277280807, + 0.05708317458629608, + 0.02114112116396427, + 0.013063421472907066, + 0.010709678754210472, + 0.023823201656341553, + 0.015152739360928535, + 0.014180583879351616, + 0.013099324889481068, + 0.05439290404319763, + 0.019888170063495636, + 0.01578272134065628, + 0.013150239363312721, + 0.025584828108549118, + 0.004053293727338314, + 0.00772838806733489, + 0.003762347623705864, + 0.010166979394853115, + 0.010445194318890572, + 0.0040227570571005344, + 0.0043115816079080105, + 0.004471697844564915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.05033217743039131 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137058138847351 + ], + "opposite_logprob_tokens": [ + -0.3870581090450287 + ], + "logprob_reward": -1.137058138847351, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137058138847351, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06644411385059357, + 0.47729602456092834, + 0.0679582953453064, + 0.17015041410923004, + 0.011002419516444206, + 0.034501004964113235, + 0.01655431091785431, + 0.047952793538570404, + 0.038534145802259445, + 0.0696064680814743 + ], + "credit_map": [ + 0.010538645088672638, + 0.008301928639411926, + 0.005997384898364544, + 0.006312924437224865, + 0.015179245732724667, + 0.005627022590488195, + 0.006354447454214096, + 0.008132511749863625, + 0.08974312990903854, + 0.061612945050001144, + 0.029374826699495316, + 0.02276831492781639, + 0.17976605892181396, + 0.03187798336148262, + 0.029465559870004654, + 0.03268720209598541, + 0.00892734993249178, + 0.009770521894097328, + 0.004801319446414709, + 0.005910180043429136, + 0.02443561889231205, + 0.004890235606580973, + 0.004703926388174295, + 0.004519145470112562, + 0.018475191667675972, + 0.01577896997332573, + 0.012017006985843182, + 0.03888271749019623, + 0.03378414362668991, + 0.016300639137625694, + 0.015412265434861183, + 0.019499488174915314, + 0.0017186388140544295, + 0.001997214276343584, + 0.0013475532177835703, + 0.001209844253025949, + 0.0015594852156937122, + 0.0009215838508680463, + 0.001186596928164363, + 0.0010615027276799083, + 0.0027620824985206127, + 0.004065924324095249, + 0.0028418872971087694, + 0.009024388156831264, + 0.006506999954581261, + 0.0030702727381139994, + 0.003765309462323785, + 0.0024641386698931456, + 0.002800519810989499, + 0.002165670972317457, + 0.0014339085901156068, + 0.002758848248049617, + 0.0020688362419605255, + 0.001473087933845818, + 0.0016343733295798302, + 0.0022190662566572428, + 0.004744793754070997, + 0.005326079670339823, + 0.0035890669096261263, + 0.015409880317747593, + 0.0057747503742575645, + 0.0037454375997185707, + 0.0036189481616020203, + 0.005743836052715778, + 0.003713758662343025, + 0.004051148891448975, + 0.0029074535705149174, + 0.011138253845274448, + 0.005187122616916895, + 0.00343346712179482, + 0.0028847402427345514, + 0.005218200851231813, + 0.006073114462196827, + 0.008680778555572033, + 0.004493341781198978, + 0.016624368727207184, + 0.008624359965324402, + 0.009203450754284859, + 0.004554554354399443, + 0.011352503672242165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -1.137058138847351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707367539405823 + ], + "opposite_logprob_tokens": [ + -2.1270737648010254 + ], + "logprob_reward": -0.12707367539405823, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707367539405823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12932032346725464, + 0.2921498119831085, + 0.07494432479143143, + 0.14317983388900757, + 0.01514044962823391, + 0.13125179708003998, + 0.06034480407834053, + 0.06577946245670319, + 0.061369121074676514, + 0.02652009017765522 + ], + "credit_map": [ + 0.01764841563999653, + 0.014608023688197136, + 0.018298711627721786, + 0.01817242056131363, + 0.022524740546941757, + 0.010935584083199501, + 0.01484919898211956, + 0.012283215299248695, + 0.054263338446617126, + 0.037704985588788986, + 0.01880587823688984, + 0.01502530463039875, + 0.10814446210861206, + 0.01899784989655018, + 0.0186283178627491, + 0.02057967334985733, + 0.01378986332565546, + 0.007175194099545479, + 0.00408836780115962, + 0.00665976433083415, + 0.0279852282255888, + 0.005027261562645435, + 0.004878545179963112, + 0.005340102594345808, + 0.02935050241649151, + 0.007651854772120714, + 0.008957098238170147, + 0.04912622645497322, + 0.02177688479423523, + 0.006873461417853832, + 0.009173338301479816, + 0.010270466096699238, + 0.00176517094951123, + 0.0027266694232821465, + 0.0013596020871773362, + 0.0021173451095819473, + 0.002406822517514229, + 0.0014468517620116472, + 0.0017712765838950872, + 0.0015467110788449645, + 0.015226454474031925, + 0.014269625768065453, + 0.012005629017949104, + 0.027572182938456535, + 0.025155480951070786, + 0.012832661159336567, + 0.013441639021039009, + 0.010748128406703472, + 0.004927239380776882, + 0.00784708559513092, + 0.004734522197395563, + 0.014891141094267368, + 0.01223763544112444, + 0.005356526467949152, + 0.005583611782640219, + 0.004767043050378561, + 0.0060514360666275024, + 0.005260740406811237, + 0.0038227643817663193, + 0.022064344957470894, + 0.008143872022628784, + 0.005594498943537474, + 0.0041966731660068035, + 0.010645133443176746, + 0.0060050454922020435, + 0.005869653541594744, + 0.004994409158825874, + 0.019478170201182365, + 0.007435559760779142, + 0.005166786722838879, + 0.005127348471432924, + 0.007292150054126978, + 0.002075495198369026, + 0.005067285615950823, + 0.0025045780930668116, + 0.004718483425676823, + 0.005641607567667961, + 0.0022889841347932816, + 0.0024741957895457745, + 0.0017494582571089268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.12707367539405823 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645249530673027 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645249530673027, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645249530673027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08875185251235962, + 0.28207260370254517, + 0.14041636884212494, + 0.1550447791814804, + 0.033940378576517105, + 0.06916768848896027, + 0.06137898191809654, + 0.03576740622520447, + 0.06464314460754395, + 0.06881679594516754 + ], + "credit_map": [ + 0.013067176565527916, + 0.01356988400220871, + 0.00942951999604702, + 0.007797681260854006, + 0.01979375071823597, + 0.00853649526834488, + 0.008676629513502121, + 0.00788071472197771, + 0.05361783131957054, + 0.035085055977106094, + 0.016918394714593887, + 0.015497947111725807, + 0.1008945181965828, + 0.019596686586737633, + 0.019562523812055588, + 0.020899655297398567, + 0.024571998044848442, + 0.014787420630455017, + 0.007573567796498537, + 0.013014713302254677, + 0.05239900201559067, + 0.01000155694782734, + 0.008772218599915504, + 0.009295901283621788, + 0.012712035328149796, + 0.01425753254443407, + 0.01153858844190836, + 0.034930143505334854, + 0.05115427076816559, + 0.009940878488123417, + 0.01128754299134016, + 0.00922378059476614, + 0.003973023965954781, + 0.005739330314099789, + 0.004296645987778902, + 0.004092778079211712, + 0.004999116528779268, + 0.0037199873477220535, + 0.0035884708631783724, + 0.003531024791300297, + 0.00704844668507576, + 0.007272565271705389, + 0.00533026410266757, + 0.020085548982024193, + 0.012413310818374157, + 0.004883382003754377, + 0.007149245589971542, + 0.004984925035387278, + 0.0061102197505533695, + 0.008418169803917408, + 0.0041409884579479694, + 0.013895077630877495, + 0.009839162230491638, + 0.006242761854082346, + 0.006220377050340176, + 0.006512220483273268, + 0.003977527841925621, + 0.004906163550913334, + 0.002798704896122217, + 0.007782557047903538, + 0.0073051368817687035, + 0.0032945594284683466, + 0.0026573939248919487, + 0.003045362187549472, + 0.006812978070229292, + 0.006538790185004473, + 0.005254892632365227, + 0.020542504265904427, + 0.00813275296241045, + 0.005333758890628815, + 0.004523023031651974, + 0.007504440378397703, + 0.006410188972949982, + 0.007989888079464436, + 0.004987532738596201, + 0.02013506181538105, + 0.00857806857675314, + 0.007312076631933451, + 0.004223247524350882, + 0.009180733934044838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.008645249530673027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot framed as a close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a canted Dutch angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the high-risk tornado zone appear as a viscous, luminous, scarlet non-Newtonian fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless fiery red droplets frozen in mid-air with trailing strands of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497878044843674 + ], + "opposite_logprob_tokens": [ + -2.929978847503662 + ], + "logprob_reward": -0.05497878044843674, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497878044843674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11197976768016815, + 0.09369124472141266, + 0.25306612253189087, + 0.20615650713443756, + 0.030743364244699478, + 0.07319270074367523, + 0.046565860509872437, + 0.07475332170724869, + 0.049581319093704224, + 0.0602698028087616 + ], + "credit_map": [ + 0.010162710212171078, + 0.022272899746894836, + 0.013980940915644169, + 0.012636220082640648, + 0.018042901530861855, + 0.011103853583335876, + 0.014163367450237274, + 0.00961688719689846, + 0.0161848496645689, + 0.012339380569756031, + 0.0058692460879683495, + 0.006312277168035507, + 0.032053254544734955, + 0.006256209686398506, + 0.007588949520140886, + 0.007087083999067545, + 0.04875751957297325, + 0.023764578625559807, + 0.014001104980707169, + 0.021336527541279793, + 0.0946728065609932, + 0.017614856362342834, + 0.014549335464835167, + 0.018369415774941444, + 0.014744308777153492, + 0.011793806217610836, + 0.008583378046751022, + 0.07607617229223251, + 0.06726787239313126, + 0.009395007975399494, + 0.01250376645475626, + 0.005792213138192892, + 0.0035699692089110613, + 0.004657285753637552, + 0.0024224764201790094, + 0.007493366487324238, + 0.00430819857865572, + 0.002633980242535472, + 0.0025042362976819277, + 0.003153851954266429, + 0.006425325758755207, + 0.008779933676123619, + 0.004945123568177223, + 0.0215590912848711, + 0.013134957291185856, + 0.0056080203503370285, + 0.007215689402073622, + 0.005524567328393459, + 0.004028211813420057, + 0.006863567046821117, + 0.0027762851677834988, + 0.016613803803920746, + 0.005485085304826498, + 0.0027560146991163492, + 0.003182817017659545, + 0.004860077518969774, + 0.006562423892319202, + 0.00627024844288826, + 0.004669218324124813, + 0.02567276731133461, + 0.010637307539582253, + 0.005895076785236597, + 0.00518587464466691, + 0.009860415011644363, + 0.004278766456991434, + 0.004443794954568148, + 0.003417543601244688, + 0.01632923074066639, + 0.00561594357714057, + 0.004403229802846909, + 0.003097082022577524, + 0.007995733059942722, + 0.004398178309202194, + 0.011551701463758945, + 0.005125020630657673, + 0.0072405473329126835, + 0.013597843237221241, + 0.0072762067429721355, + 0.0058057489804923534, + 0.005274561233818531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.05497878044843674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_tensor([9152, 9152])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_tensor([9152, 9152])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8bbedbd50945ebaed3eafb489d966a5726d5fc7a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_tensor([9152, 9152])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly dynamic overhead shot, the camera swiveling in a 360-degree arc around its subject at breakneck speed, captures the stunning spectacle of a massive Godzilla standing amidst the ruins of a shattered city. The beast, covered in rough, dark gray scales resembling the texture of ancient stone, its massive, muscular frame imposing itself against the night sky. The wreckage of skyscrapers lies like toys beneath its feet. Its head violently throws back, its massive jaws stretched to their limit. With a roar that rends the heavens, a powerful wave of energy erupts from its mouth, visually distorting the surrounding, smoky air. At the climax of this roar, the camera's rapid rotation trajectory just captured the key details: from the deepest part of its throat, a bright and dazzling atomic blue light began to pulsate and flicker. This supernatural light spread upward, illuminating the internal structure of its mouth and its dagger-sharp teeth, and illuminating its rough skin texture and the surrounding smoke and dust, casting a huge and dynamic shadow on the ruined cityscape. The whole picture is full of destructive power and an awe-inspiring doomsday atmosphere.", + "seed": 13462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13462_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1298352735582739, + "vlm_logprob": -0.1298352735582739, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.1298352735582739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005140183959156275 + ], + "opposite_logprob_tokens": [ + -5.75514030456543 + ], + "logprob_reward": -0.005140183959156275, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005140183959156275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09531158953905106, + 0.07062671333551407, + 0.08496460318565369, + 0.08550000190734863, + 0.34839823842048645, + 0.056908998638391495, + 0.08044689893722534, + 0.07383361458778381, + 0.04244253411889076, + 0.06156676635146141 + ], + "credit_map": [ + 0.016876710578799248, + 0.016251681372523308, + 0.01333436369895935, + 0.01137683354318142, + 0.009099761955440044, + 0.011681614443659782, + 0.008567776530981064, + 0.008122852072119713, + 0.01378222182393074, + 0.01724519208073616, + 0.00796249508857727, + 0.0048397257924079895, + 0.0068939984776079655, + 0.007261407095938921, + 0.00484403595328331, + 0.007797634229063988, + 0.007180711254477501, + 0.011739286594092846, + 0.008208215236663818, + 0.009099099785089493, + 0.013023548759520054, + 0.01107157114893198, + 0.012287751771509647, + 0.01235442515462637, + 0.004878157749772072, + 0.012398078106343746, + 0.00639741076156497, + 0.00798600260168314, + 0.022283006459474564, + 0.014543702825903893, + 0.010289517231285572, + 0.006724128499627113, + 0.038145553320646286, + 0.04492424055933952, + 0.036542028188705444, + 0.024468446150422096, + 0.07911086827516556, + 0.04612896218895912, + 0.04781950265169144, + 0.03125864267349243, + 0.011379540897905827, + 0.006560036912560463, + 0.005270055960863829, + 0.004956163000315428, + 0.00529082864522934, + 0.013309176079928875, + 0.005636208690702915, + 0.004506987519562244, + 0.013240419328212738, + 0.009467068128287792, + 0.008919291198253632, + 0.008706942200660706, + 0.00731505686417222, + 0.006090839393436909, + 0.01432809792459011, + 0.012379182502627373, + 0.008633652701973915, + 0.00883596669882536, + 0.011018396355211735, + 0.006159941665828228, + 0.008603563532233238, + 0.016617625951766968, + 0.007593389600515366, + 0.0063710762187838554, + 0.0032780396286398172, + 0.005086048040539026, + 0.0057466602884233, + 0.0059672691859304905, + 0.005327765829861164, + 0.004556751344352961, + 0.005466950125992298, + 0.007013047579675913, + 0.006101674400269985, + 0.00434452760964632, + 0.022728141397237778, + 0.006373005919158459, + 0.003452264005318284, + 0.011191715486347675, + 0.0035466866102069616, + 0.0038287495262920856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005140183959156275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866499826312065 + ], + "opposite_logprob_tokens": [ + -3.0486650466918945 + ], + "logprob_reward": -0.04866499826312065, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866499826312065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19189482927322388, + 0.06936861574649811, + 0.036359693855047226, + 0.13051138818264008, + 0.24311846494674683, + 0.07481788098812103, + 0.15295135974884033, + 0.030867286026477814, + 0.040193479508161545, + 0.02991701476275921 + ], + "credit_map": [ + 0.031076442450284958, + 0.026119275018572807, + 0.027125313878059387, + 0.026042906567454338, + 0.017522282898426056, + 0.02282252162694931, + 0.018242180347442627, + 0.02294391207396984, + 0.011781343258917332, + 0.015634862706065178, + 0.009455092251300812, + 0.005499221384525299, + 0.0060419924557209015, + 0.007198928389698267, + 0.00555352121591568, + 0.008203656412661076, + 0.0030303113162517548, + 0.005360796581953764, + 0.0033223479986190796, + 0.003756936639547348, + 0.005588477477431297, + 0.004795973654836416, + 0.005637009162455797, + 0.004867842420935631, + 0.02005510777235031, + 0.024980604648590088, + 0.00990641675889492, + 0.011189328506588936, + 0.01397052314132452, + 0.02078407071530819, + 0.012790088541805744, + 0.016835233196616173, + 0.02646842785179615, + 0.032866086810827255, + 0.025582706555724144, + 0.01754974015057087, + 0.05553865805268288, + 0.03239881992340088, + 0.032696425914764404, + 0.020017605274915695, + 0.016737740486860275, + 0.009007446467876434, + 0.0036670295521616936, + 0.0041152373887598515, + 0.008078217506408691, + 0.021059885621070862, + 0.00639410549774766, + 0.0057582189328968525, + 0.01751512847840786, + 0.0217742882668972, + 0.03304487466812134, + 0.025968873873353004, + 0.01141788437962532, + 0.010233879089355469, + 0.019360065460205078, + 0.013636362738907337, + 0.0036828634329140186, + 0.0037791491486132145, + 0.0038940859958529472, + 0.002839412773028016, + 0.003488323651254177, + 0.00661954702809453, + 0.003913605585694313, + 0.002650299109518528, + 0.003926063422113657, + 0.00499714445322752, + 0.007232814095914364, + 0.006244842428714037, + 0.003951284568756819, + 0.005330908112227917, + 0.004120602738112211, + 0.004389819223433733, + 0.002678357996046543, + 0.0019347588531672955, + 0.010828523896634579, + 0.0033147649373859167, + 0.0017492834012955427, + 0.004736038390547037, + 0.002431809203699231, + 0.002243479946628213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.04866499826312065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607470214366913 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607470214366913, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607470214366913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04752948507666588, + 0.0506792776286602, + 0.021299077197909355, + 0.04559493809938431, + 0.14281776547431946, + 0.06649419665336609, + 0.29588642716407776, + 0.09178987145423889, + 0.199519544839859, + 0.038389429450035095 + ], + "credit_map": [ + 0.006080338731408119, + 0.00664589274674654, + 0.007711627520620823, + 0.00817734096199274, + 0.003894572611898184, + 0.003705760231241584, + 0.006586871109902859, + 0.00472708186134696, + 0.009358469396829605, + 0.010460468009114265, + 0.007113819941878319, + 0.00439872220158577, + 0.003764136927202344, + 0.003968888893723488, + 0.0057690865360200405, + 0.005845685955137014, + 0.0019654384814202785, + 0.0033243950456380844, + 0.0018509951187297702, + 0.003002436365932226, + 0.0034956878516823053, + 0.003119453089311719, + 0.0026699367444962263, + 0.0018707358976826072, + 0.0045532407239079475, + 0.006986455526202917, + 0.0038759191520512104, + 0.0052384259179234505, + 0.006320843007415533, + 0.005806517321616411, + 0.0044752461835742, + 0.008338287472724915, + 0.015249618329107761, + 0.019008342176675797, + 0.014618752524256706, + 0.01397291012108326, + 0.02918202616274357, + 0.01774020306766033, + 0.018333520740270615, + 0.014712385833263397, + 0.013061788864433765, + 0.006828760728240013, + 0.00607994943857193, + 0.003979238215833902, + 0.008797350339591503, + 0.015044287778437138, + 0.006545430980622768, + 0.00615739356726408, + 0.06085667014122009, + 0.02709263376891613, + 0.032196059823036194, + 0.01925697736442089, + 0.023011280223727226, + 0.019215986132621765, + 0.05116385221481323, + 0.06309295445680618, + 0.006664613727480173, + 0.012935958802700043, + 0.013681154698133469, + 0.007553713396191597, + 0.015252112410962582, + 0.013251738622784615, + 0.012117328122258186, + 0.010333257727324963, + 0.01756734400987625, + 0.03567097708582878, + 0.019316043704748154, + 0.025570906698703766, + 0.025828134268522263, + 0.02158258855342865, + 0.021618088707327843, + 0.032365456223487854, + 0.0036051454953849316, + 0.004344146233052015, + 0.004622821696102619, + 0.0056036957539618015, + 0.006624160334467888, + 0.003968984354287386, + 0.005000133067369461, + 0.004620338790118694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.01607470214366913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319589558057487 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.00319589558057487, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319589558057487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08443458378314972, + 0.0446326807141304, + 0.05196818709373474, + 0.039411723613739014, + 0.12324444204568863, + 0.1335451453924179, + 0.15553203225135803, + 0.043973758816719055, + 0.13970635831356049, + 0.1835511028766632 + ], + "credit_map": [ + 0.011644365265965462, + 0.010775652714073658, + 0.01192477997392416, + 0.016571851447224617, + 0.00846916250884533, + 0.006415176670998335, + 0.011667822487652302, + 0.006965773645788431, + 0.010392936877906322, + 0.00908737163990736, + 0.006695077754557133, + 0.0030861664563417435, + 0.003903290256857872, + 0.005058065988123417, + 0.0026449449360370636, + 0.003764827735722065, + 0.0047287531197071075, + 0.007470411714166403, + 0.004543445073068142, + 0.007542906794697046, + 0.007659457623958588, + 0.005815343931317329, + 0.008200112730264664, + 0.006007756106555462, + 0.005061131902039051, + 0.004847193136811256, + 0.0031486358493566513, + 0.0043357862159609795, + 0.0056267715990543365, + 0.005123565439134836, + 0.004223690368235111, + 0.007044951431453228, + 0.014462003484368324, + 0.017274057492613792, + 0.013798505067825317, + 0.008385267108678818, + 0.025741031393408775, + 0.016426442191004753, + 0.016775676980614662, + 0.010381458327174187, + 0.025355882942676544, + 0.019228821620345116, + 0.010954740457236767, + 0.008782441727817059, + 0.01851857826113701, + 0.02888786792755127, + 0.013376531191170216, + 0.008440276607871056, + 0.023191560059785843, + 0.027014490216970444, + 0.019964173436164856, + 0.022136731073260307, + 0.014465843327343464, + 0.00926829781383276, + 0.01964833214879036, + 0.019842613488435745, + 0.004306678660213947, + 0.005468298681080341, + 0.006947013083845377, + 0.005278289318084717, + 0.006401785183697939, + 0.00548264617100358, + 0.005506324581801891, + 0.0045827217400074005, + 0.010758024640381336, + 0.025644956156611443, + 0.01696735993027687, + 0.02192293107509613, + 0.01743955910205841, + 0.0153846750035882, + 0.016657007858157158, + 0.014931854791939259, + 0.018711835145950317, + 0.01330091618001461, + 0.055027786642313004, + 0.02083364874124527, + 0.012533411383628845, + 0.02898423559963703, + 0.01923629455268383, + 0.01492297276854515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.00319589558057487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761005878448486 + ], + "opposite_logprob_tokens": [ + -0.8261005878448486 + ], + "logprob_reward": -0.5761005878448486, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761005878448486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054193396121263504, + 0.08698893338441849, + 0.05054905638098717, + 0.040780432522296906, + 0.1058306023478508, + 0.21805085241794586, + 0.13582845032215118, + 0.12028305232524872, + 0.1166597232222557, + 0.07083550095558167 + ], + "credit_map": [ + 0.008115950971841812, + 0.006254543084651232, + 0.00864364579319954, + 0.008083340711891651, + 0.005625633057206869, + 0.005345924291759729, + 0.00598511565476656, + 0.0061392406933009624, + 0.016849366948008537, + 0.020263653248548508, + 0.010975445620715618, + 0.006948146503418684, + 0.007269378751516342, + 0.009050838649272919, + 0.007516545709222555, + 0.008115561679005623, + 0.004326772876083851, + 0.0071738241240382195, + 0.0035151904448866844, + 0.006431375630199909, + 0.011234203353524208, + 0.005506514571607113, + 0.008220923133194447, + 0.004140250850468874, + 0.0038688473869115114, + 0.008613700978457928, + 0.003136090701445937, + 0.0033528958447277546, + 0.007223514840006828, + 0.007544818799942732, + 0.0033767546992748976, + 0.0036638102028518915, + 0.015534523874521255, + 0.01661841757595539, + 0.00957155879586935, + 0.011327249929308891, + 0.01591162383556366, + 0.013367118313908577, + 0.014713581651449203, + 0.008786534890532494, + 0.04601966217160225, + 0.029374904930591583, + 0.01516040787100792, + 0.014972012490034103, + 0.021105017513036728, + 0.05702038109302521, + 0.01922757364809513, + 0.015170885249972343, + 0.02561338059604168, + 0.013434319756925106, + 0.016606682911515236, + 0.013270694762468338, + 0.011317181400954723, + 0.009493252262473106, + 0.024636680260300636, + 0.021456250920891762, + 0.01312598492950201, + 0.016424698755145073, + 0.015255454927682877, + 0.010127706453204155, + 0.014599824324250221, + 0.02517426200211048, + 0.014287748374044895, + 0.011287369765341282, + 0.00939831044524908, + 0.02289586327970028, + 0.013870804570615292, + 0.014285353012382984, + 0.014271870255470276, + 0.013624920509755611, + 0.010159436613321304, + 0.018153171986341476, + 0.006217270623892546, + 0.006256671156734228, + 0.018796158954501152, + 0.00891382060945034, + 0.006188430357724428, + 0.010663095861673355, + 0.006433296948671341, + 0.007366756442934275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.5761005878448486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic overhead shot, the camera swiveling in a 360-degree arc around its subject at breakneck speed, captures the stunning spectacle of a massive Godzilla standing amidst the ruins of a shattered city. The beast, covered in rough, dark gray scales resembling the texture of ancient stone, its massive, muscular frame imposing itself against the night sky. The wreckage of skyscrapers lies like toys beneath its feet. Its head violently throws back, its massive jaws stretched to their limit. With a roar that rends the heavens, a powerful wave of energy erupts from its mouth, visually distorting the surrounding, smoky air. At the climax of this roar, the camera's rapid rotation trajectory just captured the key details: from the deepest part of its throat, a bright and dazzling atomic blue light began to pulsate and flicker. This supernatural light spread upward, illuminating the internal structure of its mouth and its dagger-sharp teeth, and illuminating its rough skin texture and the surrounding smoke and dust, casting a huge and dynamic shadow on the ruined cityscape. The whole picture is full of destructive power and an awe-inspiring doomsday atmosphere.", + "seed": 13463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13463_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.035621865349821744, + "vlm_logprob": -0.035621865349821744, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.035621865349821744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037782251834869385 + ], + "opposite_logprob_tokens": [ + -3.4127821922302246 + ], + "logprob_reward": -0.037782251834869385, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037782251834869385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20136402547359467, + 0.041926417499780655, + 0.10766950249671936, + 0.09593326598405838, + 0.06445132195949554, + 0.20616468787193298, + 0.13907533884048462, + 0.05262531340122223, + 0.03269879147410393, + 0.05809137225151062 + ], + "credit_map": [ + 0.01679125800728798, + 0.026341063901782036, + 0.02708139456808567, + 0.02018655836582184, + 0.02391938865184784, + 0.03022041730582714, + 0.02199561893939972, + 0.034828316420316696, + 0.004699208773672581, + 0.008524552918970585, + 0.008328038267791271, + 0.004712574649602175, + 0.0028338038828223944, + 0.0038313919212669134, + 0.0047270855866372585, + 0.004269761964678764, + 0.011193802580237389, + 0.012900829315185547, + 0.0162919033318758, + 0.01689673401415348, + 0.012578235007822514, + 0.01285985205322504, + 0.011226043105125427, + 0.013722101226449013, + 0.015053692273795605, + 0.011671526357531548, + 0.010047446005046368, + 0.012997745536267757, + 0.010831915773451328, + 0.014341534115374088, + 0.010787365026772022, + 0.010202035307884216, + 0.010572553612291813, + 0.006915028672665358, + 0.006236920598894358, + 0.007561543490737677, + 0.006596708204597235, + 0.008656596764922142, + 0.010098740458488464, + 0.007813229225575924, + 0.028700949624180794, + 0.0733487606048584, + 0.006085759494453669, + 0.010486054234206676, + 0.00997764989733696, + 0.05726686120033264, + 0.01019430160522461, + 0.01010433305054903, + 0.02039986662566662, + 0.01240554265677929, + 0.01656300202012062, + 0.01330993976444006, + 0.01905266009271145, + 0.01534311193972826, + 0.024248894304037094, + 0.01775231771171093, + 0.00568069564178586, + 0.003850095672532916, + 0.005629676394164562, + 0.008706284686923027, + 0.005353902466595173, + 0.006479651667177677, + 0.011242113076150417, + 0.0056828949600458145, + 0.004294483922421932, + 0.0035783869680017233, + 0.006133537273854017, + 0.0043736048974096775, + 0.0037496546283364296, + 0.0029017534106969833, + 0.004489061888307333, + 0.003178305458277464, + 0.006504475604742765, + 0.004362863954156637, + 0.008992848917841911, + 0.014396625570952892, + 0.004687799606472254, + 0.007085018325597048, + 0.005928344093263149, + 0.006133395712822676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.037782251834869385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982505038380623 + ], + "opposite_logprob_tokens": [ + -3.52982497215271 + ], + "logprob_reward": -0.02982505038380623, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982505038380623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13134464621543884, + 0.1514192670583725, + 0.08551483601331711, + 0.048279449343681335, + 0.044094979763031006, + 0.19925977289676666, + 0.08440615236759186, + 0.06583073735237122, + 0.10172621160745621, + 0.0881238654255867 + ], + "credit_map": [ + 0.012924518436193466, + 0.0208086259663105, + 0.020225390791893005, + 0.013538109138607979, + 0.014179769903421402, + 0.019018372520804405, + 0.013412533327937126, + 0.017237327992916107, + 0.020143695175647736, + 0.03479725867509842, + 0.03910071775317192, + 0.0161470677703619, + 0.010085695423185825, + 0.00941134337335825, + 0.011232387274503708, + 0.010501095093786716, + 0.011725974269211292, + 0.011980998329818249, + 0.010161678306758404, + 0.01267620362341404, + 0.008433009497821331, + 0.012048687785863876, + 0.009676565416157246, + 0.008811723440885544, + 0.007635973859578371, + 0.0042457012459635735, + 0.0056541794911026955, + 0.006862008012831211, + 0.005303812213242054, + 0.006758478935807943, + 0.006323806010186672, + 0.005495491437613964, + 0.00568985752761364, + 0.005908366292715073, + 0.006446418352425098, + 0.004816581029444933, + 0.004670793190598488, + 0.005127243231981993, + 0.006831017788499594, + 0.004604702815413475, + 0.03033578023314476, + 0.07234539091587067, + 0.0053778160363435745, + 0.00935965683311224, + 0.008410736918449402, + 0.056695032864809036, + 0.008726637810468674, + 0.008008736185729504, + 0.009767282754182816, + 0.011269202455878258, + 0.009711332619190216, + 0.01130806002765894, + 0.008120136335492134, + 0.013618653640151024, + 0.011308236978948116, + 0.00930324662476778, + 0.006776574533432722, + 0.0033480494748800993, + 0.0061411927454173565, + 0.010725821368396282, + 0.0041263168677687645, + 0.0060813697054982185, + 0.022390367463231087, + 0.00624104356393218, + 0.014969763346016407, + 0.010990990325808525, + 0.022703835740685463, + 0.008054367266595364, + 0.013583788648247719, + 0.008764478377997875, + 0.010220838710665703, + 0.012438150122761726, + 0.013324410654604435, + 0.014952542260289192, + 0.012462694197893143, + 0.020126745104789734, + 0.005183800123631954, + 0.0067520905286073685, + 0.007525584194809198, + 0.007795992307364941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.02982505038380623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819844290614128 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819844290614128, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819844290614128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14192160964012146, + 0.0925680622458458, + 0.1502579152584076, + 0.04245563596487045, + 0.027695246040821075, + 0.15357324481010437, + 0.06465847790241241, + 0.08794283121824265, + 0.07007650285959244, + 0.16885049641132355 + ], + "credit_map": [ + 0.014543971046805382, + 0.017380187287926674, + 0.0183542650192976, + 0.015764180570840836, + 0.021044978871941566, + 0.01625111699104309, + 0.019435735419392586, + 0.019147170707583427, + 0.011408966034650803, + 0.015109367668628693, + 0.02140670083463192, + 0.009640214964747429, + 0.008463763631880283, + 0.012210057117044926, + 0.0064790998585522175, + 0.007849890738725662, + 0.01978665217757225, + 0.020602934062480927, + 0.016969436779618263, + 0.025140026584267616, + 0.023942217230796814, + 0.015041877515614033, + 0.015341284684836864, + 0.013433486223220825, + 0.00745572242885828, + 0.005495558492839336, + 0.003915103618055582, + 0.0066511244513094425, + 0.0025821756571531296, + 0.006666939239948988, + 0.00455324724316597, + 0.005135764833539724, + 0.0026629019994288683, + 0.0027386017609387636, + 0.0031842130701988935, + 0.004858916625380516, + 0.0034413584508001804, + 0.0034495703876018524, + 0.005525956396013498, + 0.0018337273504585028, + 0.018698163330554962, + 0.05448411405086517, + 0.004682143684476614, + 0.008204786106944084, + 0.006446176208555698, + 0.04883246123790741, + 0.0062035731971263885, + 0.0060218265280127525, + 0.008557799272239208, + 0.006779085844755173, + 0.010420518927276134, + 0.006008767988532782, + 0.006906295195221901, + 0.0075450558215379715, + 0.012122249230742455, + 0.0063187056221067905, + 0.0090487627312541, + 0.004197185859084129, + 0.007790030911564827, + 0.015601511113345623, + 0.005296130198985338, + 0.005442880559712648, + 0.03316164016723633, + 0.007404693868011236, + 0.008671553805470467, + 0.007696795277297497, + 0.006674605887383223, + 0.007955120876431465, + 0.009438103064894676, + 0.007504167966544628, + 0.015415268950164318, + 0.006720888428390026, + 0.014569257386028767, + 0.02147042751312256, + 0.02798095904290676, + 0.035741176456213, + 0.015772191807627678, + 0.023951629176735878, + 0.01633470691740513, + 0.013030142523348331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.029819844290614128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017175221582874656 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017175221582874656, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017175221582874656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2691720426082611, + 0.04028157889842987, + 0.03011094406247139, + 0.059037186205387115, + 0.04913926497101784, + 0.12567628920078278, + 0.08405863493680954, + 0.05460294336080551, + 0.052326515316963196, + 0.23559460043907166 + ], + "credit_map": [ + 0.02465023286640644, + 0.030435388907790184, + 0.03379005938768387, + 0.02830037847161293, + 0.04070187732577324, + 0.02747984044253826, + 0.039577554911375046, + 0.044236745685338974, + 0.00564139150083065, + 0.007744180969893932, + 0.0063176387920975685, + 0.005571630317717791, + 0.0028815397527068853, + 0.004624943248927593, + 0.0031033712439239025, + 0.004396885167807341, + 0.004456734284758568, + 0.0038069849833846092, + 0.0032236722763627768, + 0.005142483860254288, + 0.004810215439647436, + 0.0030563734471797943, + 0.0030243825167417526, + 0.0025901005137711763, + 0.0106098223477602, + 0.00513334758579731, + 0.005890177562832832, + 0.009940839372575283, + 0.007346014492213726, + 0.006433432921767235, + 0.007638974580913782, + 0.0060445815324783325, + 0.00673929275944829, + 0.004413996823132038, + 0.0065260534174740314, + 0.006734595634043217, + 0.00467207795009017, + 0.006237059831619263, + 0.008376289159059525, + 0.005439900327473879, + 0.015360165387392044, + 0.04376785829663277, + 0.004024127032607794, + 0.006024284288287163, + 0.005517740733921528, + 0.04015696421265602, + 0.0048479656688869, + 0.005977186840027571, + 0.011982467956840992, + 0.007640181574970484, + 0.009713406674563885, + 0.007913553155958652, + 0.012039098888635635, + 0.009466932155191898, + 0.014714674092829227, + 0.010588320903480053, + 0.005620499141514301, + 0.0029486126732081175, + 0.005753334145992994, + 0.00881169829517603, + 0.0035663133021444082, + 0.004426124971359968, + 0.017865575850009918, + 0.005610790103673935, + 0.006536457221955061, + 0.007248150184750557, + 0.006589727941900492, + 0.007466483395546675, + 0.005607318133115768, + 0.003946786746382713, + 0.010501543059945107, + 0.004430050030350685, + 0.026852501556277275, + 0.03624095395207405, + 0.04595614969730377, + 0.035182077437639236, + 0.01841582916676998, + 0.03277727589011192, + 0.022115418687462807, + 0.018054410815238953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0017175221582874656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896465808153152 + ], + "opposite_logprob_tokens": [ + -2.5789647102355957 + ], + "logprob_reward": -0.07896465808153152, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896465808153152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26065826416015625, + 0.04783645272254944, + 0.09043706208467484, + 0.0482945553958416, + 0.015839537605643272, + 0.13878102600574493, + 0.04560455307364464, + 0.06662055850028992, + 0.07909729331731796, + 0.2068306803703308 + ], + "credit_map": [ + 0.024998163804411888, + 0.027643339708447456, + 0.034335363656282425, + 0.029590245336294174, + 0.03804182633757591, + 0.027842681854963303, + 0.04033243656158447, + 0.03787422925233841, + 0.006149929482489824, + 0.005979278590530157, + 0.005057217087596655, + 0.006778277922421694, + 0.003774345386773348, + 0.006155775394290686, + 0.007916925475001335, + 0.006024699658155441, + 0.01050865463912487, + 0.010516275651752949, + 0.013954263180494308, + 0.014185511507093906, + 0.010382425040006638, + 0.008746957406401634, + 0.009863968007266521, + 0.012279010377824306, + 0.005575172603130341, + 0.004170305095613003, + 0.004749138839542866, + 0.005407019518315792, + 0.00442158617079258, + 0.009689449332654476, + 0.005525354295969009, + 0.008756531402468681, + 0.002113450551405549, + 0.002101539634168148, + 0.0018657274777069688, + 0.0021599363535642624, + 0.0017133656656369567, + 0.002152145141735673, + 0.0024915114045143127, + 0.0012418621918186545, + 0.01518299151211977, + 0.04795777425169945, + 0.0045238458551466465, + 0.007013155147433281, + 0.006281430833041668, + 0.04585269093513489, + 0.0055316961370408535, + 0.0064374301582574844, + 0.005333544220775366, + 0.00590620469301939, + 0.0038855646271258593, + 0.004939679987728596, + 0.004826871212571859, + 0.007069433573633432, + 0.006595452316105366, + 0.007047801744192839, + 0.006222170311957598, + 0.007372433319687843, + 0.008054687641561031, + 0.008131437003612518, + 0.008514097891747952, + 0.009286846034228802, + 0.009168507531285286, + 0.009870375506579876, + 0.009434198029339314, + 0.010258198715746403, + 0.012080753222107887, + 0.014193159528076649, + 0.009226162917912006, + 0.007946855388581753, + 0.00874344352632761, + 0.0072145164012908936, + 0.029355080798268318, + 0.029857885092496872, + 0.02896595187485218, + 0.029418742284178734, + 0.01958726905286312, + 0.034848760813474655, + 0.01766209490597248, + 0.017134889960289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.07896465808153152 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic overhead shot, the camera swiveling in a 360-degree arc around its subject at breakneck speed, captures the stunning spectacle of a massive Godzilla standing amidst the ruins of a shattered city. The beast, covered in rough, dark gray scales resembling the texture of ancient stone, its massive, muscular frame imposing itself against the night sky. The wreckage of skyscrapers lies like toys beneath its feet. Its head violently throws back, its massive jaws stretched to their limit. With a roar that rends the heavens, a powerful wave of energy erupts from its mouth, visually distorting the surrounding, smoky air. At the climax of this roar, the camera's rapid rotation trajectory just captured the key details: from the deepest part of its throat, a bright and dazzling atomic blue light began to pulsate and flicker. This supernatural light spread upward, illuminating the internal structure of its mouth and its dagger-sharp teeth, and illuminating its rough skin texture and the surrounding smoke and dust, casting a huge and dynamic shadow on the ruined cityscape. The whole picture is full of destructive power and an awe-inspiring doomsday atmosphere.", + "seed": 13464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13464_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07027418143115938, + "vlm_logprob": -0.07027418143115938, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.07027418143115938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011399375274777412 + ], + "opposite_logprob_tokens": [ + -5.011399269104004 + ], + "logprob_reward": -0.011399375274777412, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011399375274777412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24036350846290588, + 0.13834960758686066, + 0.04952564463019371, + 0.050669148564338684, + 0.03953922539949417, + 0.1594885289669037, + 0.14657248556613922, + 0.06802342087030411, + 0.06042353808879852, + 0.04704481363296509 + ], + "credit_map": [ + 0.018001576885581017, + 0.02885940857231617, + 0.028530040755867958, + 0.04508158564567566, + 0.03108355402946472, + 0.034937694668769836, + 0.029475577175617218, + 0.024394096806645393, + 0.014728529378771782, + 0.0110773965716362, + 0.01661890558898449, + 0.025786060839891434, + 0.019839441403746605, + 0.012141761370003223, + 0.016290241852402687, + 0.021867278963327408, + 0.012995953671634197, + 0.005476380232721567, + 0.004523693583905697, + 0.004249687306582928, + 0.00604770053178072, + 0.0056182872503995895, + 0.0035304920747876167, + 0.007083449047058821, + 0.005812432151287794, + 0.004626864101737738, + 0.008301670663058758, + 0.004723664373159409, + 0.006550541613250971, + 0.010950431227684021, + 0.0036915636155754328, + 0.006011985708028078, + 0.004611704032868147, + 0.003063017036765814, + 0.0044151367619633675, + 0.00605052150785923, + 0.0063538579270243645, + 0.005946883000433445, + 0.003012605244293809, + 0.006085502449423075, + 0.01377443503588438, + 0.014692679047584534, + 0.020765449851751328, + 0.03089262545108795, + 0.02737678773701191, + 0.022314276546239853, + 0.01572304032742977, + 0.013949244283139706, + 0.021979067474603653, + 0.018591202795505524, + 0.016426363959908485, + 0.024119995534420013, + 0.014062481001019478, + 0.02046164683997631, + 0.012880361638963223, + 0.01805136725306511, + 0.011343657970428467, + 0.008147143758833408, + 0.0098731005564332, + 0.009848532266914845, + 0.004646603483706713, + 0.004964598920196295, + 0.00729708606377244, + 0.011902706697583199, + 0.011665615253150463, + 0.005772199481725693, + 0.006068983115255833, + 0.004465456120669842, + 0.007415010593831539, + 0.010448035784065723, + 0.004932606127113104, + 0.009655635803937912, + 0.009051027707755566, + 0.004210572689771652, + 0.005659190937876701, + 0.004411234520375729, + 0.006350883282721043, + 0.004167365841567516, + 0.0029968770686537027, + 0.010197660885751247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011399375274777412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819265007972717 + ], + "opposite_logprob_tokens": [ + -3.5298192501068115 + ], + "logprob_reward": -0.029819265007972717, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819265007972717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21646665036678314, + 0.16520020365715027, + 0.02838863432407379, + 0.32049649953842163, + 0.03193036839365959, + 0.04462811350822449, + 0.08500494062900543, + 0.018063364550471306, + 0.0342387855052948, + 0.05558246374130249 + ], + "credit_map": [ + 0.014305840246379375, + 0.01424705795943737, + 0.01575036719441414, + 0.018032170832157135, + 0.010330505669116974, + 0.07630537450313568, + 0.008307086303830147, + 0.05918823927640915, + 0.01568010076880455, + 0.020213516429066658, + 0.017532233148813248, + 0.026436425745487213, + 0.02867807447910309, + 0.014596977271139622, + 0.018599754199385643, + 0.023463107645511627, + 0.003477954538539052, + 0.0036066025495529175, + 0.003106264863163233, + 0.0036750410217791796, + 0.003579929005354643, + 0.0048569259233772755, + 0.0022905534133315086, + 0.003795360680669546, + 0.030699312686920166, + 0.024267785251140594, + 0.06705973297357559, + 0.025945717468857765, + 0.023911071941256523, + 0.09659089148044586, + 0.017373591661453247, + 0.0346483550965786, + 0.0036695790477097034, + 0.0018463411834090948, + 0.005798651836812496, + 0.004085069056600332, + 0.004240856505930424, + 0.0036478901747614145, + 0.0029896863270550966, + 0.005652292165905237, + 0.003982486668974161, + 0.003465144196525216, + 0.004943829961121082, + 0.004437352530658245, + 0.010244383476674557, + 0.006019801367074251, + 0.00499864062294364, + 0.006536472123116255, + 0.014379247091710567, + 0.008927937597036362, + 0.007361133117228746, + 0.010648222640156746, + 0.00901463907212019, + 0.01453404687345028, + 0.006739232689142227, + 0.01340047549456358, + 0.002624537330120802, + 0.0023671973031014204, + 0.00246685603633523, + 0.002131209010258317, + 0.0024418814573436975, + 0.0024616869632154703, + 0.0015337330987676978, + 0.002036262536421418, + 0.004201341886073351, + 0.003160264575853944, + 0.004531102254986763, + 0.0029024428222328424, + 0.005233381874859333, + 0.005866363178938627, + 0.004422248341143131, + 0.0039216396398842335, + 0.011624774895608425, + 0.004426710307598114, + 0.004657600075006485, + 0.0034844637848436832, + 0.012287653982639313, + 0.007540693040937185, + 0.0037062482442706823, + 0.007854314520955086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.029819265007972717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358619332313538 + ], + "opposite_logprob_tokens": [ + -3.6513586044311523 + ], + "logprob_reward": -0.026358619332313538, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358619332313538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4124501049518585, + 0.0823388323187828, + 0.048133429139852524, + 0.03348197415471077, + 0.021170828491449356, + 0.05276012420654297, + 0.05514845997095108, + 0.06236053258180618, + 0.15733195841312408, + 0.0748237669467926 + ], + "credit_map": [ + 0.017781168222427368, + 0.014849667437374592, + 0.027184825390577316, + 0.03766471520066261, + 0.0190155990421772, + 0.15494918823242188, + 0.02294968068599701, + 0.11805524677038193, + 0.005506086628884077, + 0.005669055040925741, + 0.005914207547903061, + 0.022635987028479576, + 0.01068361196666956, + 0.007582302205264568, + 0.010328315198421478, + 0.01401926763355732, + 0.0055266618728637695, + 0.004332058131694794, + 0.004696069750934839, + 0.006181975360959768, + 0.005382188130170107, + 0.009634644724428654, + 0.005331807304173708, + 0.007048022001981735, + 0.00354811642318964, + 0.0030906633473932743, + 0.004259069450199604, + 0.003756578080356121, + 0.008472900837659836, + 0.004970903508365154, + 0.002757501555606723, + 0.00262624048627913, + 0.0026298710145056248, + 0.0018101900350302458, + 0.0025697245728224516, + 0.002897682599723339, + 0.0032403315417468548, + 0.0027080909349024296, + 0.002180938608944416, + 0.003133997553959489, + 0.004906615242362022, + 0.0051447488367557526, + 0.006708741188049316, + 0.008012421429157257, + 0.008562724106013775, + 0.00744888698682189, + 0.006227664649486542, + 0.005748322233557701, + 0.009145808406174183, + 0.005640619434416294, + 0.006639584898948669, + 0.006279345601797104, + 0.004674910102039576, + 0.008441602811217308, + 0.005282232072204351, + 0.009044361300766468, + 0.009161676280200481, + 0.007590705994516611, + 0.007683159783482552, + 0.007155536208301783, + 0.00933636911213398, + 0.007533046416938305, + 0.005972866900265217, + 0.007927174679934978, + 0.024336444213986397, + 0.017950551584362984, + 0.019490230828523636, + 0.013830788433551788, + 0.018351906910538673, + 0.026110554113984108, + 0.015143352560698986, + 0.022118132561445236, + 0.022365402430295944, + 0.005435143131762743, + 0.006115424912422895, + 0.003910279367119074, + 0.018005866557359695, + 0.008746619336307049, + 0.004356846679002047, + 0.005888182204216719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.026358619332313538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002492894185706973 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.002492894185706973, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002492894185706973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24058076739311218, + 0.09737718105316162, + 0.02506932243704796, + 0.04287470877170563, + 0.04012925177812576, + 0.06038777530193329, + 0.048979438841342926, + 0.049348171800374985, + 0.09202113747596741, + 0.30323219299316406 + ], + "credit_map": [ + 0.012421375140547752, + 0.012741448357701302, + 0.01626509241759777, + 0.02530466392636299, + 0.016748931258916855, + 0.0846276730298996, + 0.0170229934155941, + 0.05544860661029816, + 0.011585385538637638, + 0.010726198554039001, + 0.011501879431307316, + 0.013023518025875092, + 0.01741882972419262, + 0.010548345744609833, + 0.010766636580228806, + 0.011806393042206764, + 0.0065418509766459465, + 0.0024122821632772684, + 0.00251778238452971, + 0.0021622800268232822, + 0.0024090618826448917, + 0.0032423443626612425, + 0.0018584433710202575, + 0.00392528111115098, + 0.004218469839543104, + 0.0038488004356622696, + 0.009213902987539768, + 0.004195380490273237, + 0.0046107093803584576, + 0.011844925582408905, + 0.002119373297318816, + 0.0028231542091816664, + 0.004918727092444897, + 0.0021124035120010376, + 0.008396117016673088, + 0.003897768445312977, + 0.004621095955371857, + 0.0042333295568823814, + 0.004988227970898151, + 0.006961588282138109, + 0.005990272853523493, + 0.005531494505703449, + 0.006574713625013828, + 0.008622178807854652, + 0.010275288484990597, + 0.009086607955396175, + 0.008495734073221684, + 0.005811495706439018, + 0.007147423457354307, + 0.007439584005624056, + 0.005133555270731449, + 0.005417739041149616, + 0.005589641164988279, + 0.0071535599417984486, + 0.003945505246520042, + 0.007152438163757324, + 0.0068283104337751865, + 0.005297012161463499, + 0.004031038843095303, + 0.003906600177288055, + 0.00802367553114891, + 0.010222667828202248, + 0.003650458063930273, + 0.007388412021100521, + 0.011804159730672836, + 0.011611955240368843, + 0.008615666069090366, + 0.007725979667156935, + 0.01760740391910076, + 0.016315879300236702, + 0.006437012925744057, + 0.01190309040248394, + 0.06730562448501587, + 0.01873576082289219, + 0.016422370448708534, + 0.008047286421060562, + 0.0866880789399147, + 0.02802860550582409, + 0.014853224158287048, + 0.0631512776017189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.002492894185706973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130075335502625 + ], + "opposite_logprob_tokens": [ + -1.4063007831573486 + ], + "logprob_reward": -0.28130075335502625, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130075335502625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45963963866233826, + 0.09957139939069748, + 0.031771011650562286, + 0.038943175226449966, + 0.02401631698012352, + 0.04874454066157341, + 0.13669565320014954, + 0.0388861782848835, + 0.04526462405920029, + 0.07646743208169937 + ], + "credit_map": [ + 0.020103178918361664, + 0.016819195821881294, + 0.023873623460531235, + 0.0343591570854187, + 0.02660571224987507, + 0.1831689476966858, + 0.024750476703047752, + 0.1299593448638916, + 0.013063471764326096, + 0.013372344896197319, + 0.011590437963604927, + 0.010824726894497871, + 0.015827378258109093, + 0.013116639107465744, + 0.010059657506644726, + 0.011716737411916256, + 0.006319774314761162, + 0.004303973168134689, + 0.00346092926338315, + 0.003302366007119417, + 0.0031109724659472704, + 0.00458423001691699, + 0.00250146072357893, + 0.004187303129583597, + 0.004709916654974222, + 0.002840039785951376, + 0.004909337963908911, + 0.005474031902849674, + 0.004668045323342085, + 0.008411560207605362, + 0.0021173078566789627, + 0.00581293785944581, + 0.0031661330722272396, + 0.0023539664689451456, + 0.003629853716120124, + 0.00410892628133297, + 0.0033301692456007004, + 0.002959876786917448, + 0.0018275780603289604, + 0.0026398119516670704, + 0.0041312528774142265, + 0.004203300457447767, + 0.00663306750357151, + 0.006329115014523268, + 0.008990581147372723, + 0.006652910262346268, + 0.004630233161151409, + 0.007174079306423664, + 0.023066183552145958, + 0.01315721869468689, + 0.012341360561549664, + 0.015161935240030289, + 0.014986575581133366, + 0.022956937551498413, + 0.012987720780074596, + 0.022037718445062637, + 0.0055815111845731735, + 0.004690556321293116, + 0.004956994671374559, + 0.006372310686856508, + 0.004244350828230381, + 0.0030008982867002487, + 0.004401735030114651, + 0.005637821275740862, + 0.006020492874085903, + 0.006314355880022049, + 0.006275816820561886, + 0.004200397524982691, + 0.006947639863938093, + 0.0051865968853235245, + 0.004635620396584272, + 0.0056837014853954315, + 0.01686616614460945, + 0.0055149695836007595, + 0.0055182273499667645, + 0.004589658696204424, + 0.019102638587355614, + 0.008193385787308216, + 0.004015403799712658, + 0.012666981667280197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.28130075335502625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic overhead shot, the camera swiveling in a 360-degree arc around its subject at breakneck speed, captures the stunning spectacle of a massive Godzilla standing amidst the ruins of a shattered city. The beast, covered in rough, dark gray scales resembling the texture of ancient stone, its massive, muscular frame imposing itself against the night sky. The wreckage of skyscrapers lies like toys beneath its feet. Its head violently throws back, its massive jaws stretched to their limit. With a roar that rends the heavens, a powerful wave of energy erupts from its mouth, visually distorting the surrounding, smoky air. At the climax of this roar, the camera's rapid rotation trajectory just captured the key details: from the deepest part of its throat, a bright and dazzling atomic blue light began to pulsate and flicker. This supernatural light spread upward, illuminating the internal structure of its mouth and its dagger-sharp teeth, and illuminating its rough skin texture and the surrounding smoke and dust, casting a huge and dynamic shadow on the ruined cityscape. The whole picture is full of destructive power and an awe-inspiring doomsday atmosphere.", + "seed": 13465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_13/index_9152_seed_13465_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07481651026755572, + "vlm_logprob": -0.07481651026755572, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07481651026755572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006591245997697115 + ], + "opposite_logprob_tokens": [ + -6.006591320037842 + ], + "logprob_reward": -0.006591245997697115, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006591245997697115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16336248815059662, + 0.06468416750431061, + 0.12797696888446808, + 0.0511028915643692, + 0.06487995386123657, + 0.08917343616485596, + 0.03597290813922882, + 0.07152175903320312, + 0.24468384683132172, + 0.08664161711931229 + ], + "credit_map": [ + 0.025704005733132362, + 0.024414559826254845, + 0.014138573780655861, + 0.030249468982219696, + 0.01417588908225298, + 0.022685902193188667, + 0.01464158110320568, + 0.01735251396894455, + 0.007777772843837738, + 0.011578368954360485, + 0.007662526797503233, + 0.009305739775300026, + 0.008105339482426643, + 0.007125063333660364, + 0.0062631419859826565, + 0.006866214796900749, + 0.01492650993168354, + 0.01915980689227581, + 0.019699716940522194, + 0.01412695087492466, + 0.009679495356976986, + 0.012143783271312714, + 0.02111099846661091, + 0.01712970808148384, + 0.004960952792316675, + 0.005336218513548374, + 0.007370347622781992, + 0.004593000281602144, + 0.009874355047941208, + 0.0055817654356360435, + 0.006927554961293936, + 0.006458691321313381, + 0.006494348403066397, + 0.0060784053057432175, + 0.007085171062499285, + 0.011912425979971886, + 0.0077034467831254005, + 0.007423711009323597, + 0.010116146877408028, + 0.008066300302743912, + 0.013795849867165089, + 0.007622599136084318, + 0.016543027013540268, + 0.011172824539244175, + 0.00863174069672823, + 0.009726285934448242, + 0.01234622672200203, + 0.009334883652627468, + 0.005022784229367971, + 0.003329083789139986, + 0.003827071748673916, + 0.004894162528216839, + 0.0036146240308880806, + 0.005530663765966892, + 0.0054910434409976006, + 0.004263472277671099, + 0.008025268092751503, + 0.006626671180129051, + 0.005651837680488825, + 0.015635687857866287, + 0.009422991424798965, + 0.006720765493810177, + 0.01103213056921959, + 0.008406409062445164, + 0.09000905603170395, + 0.009766347706317902, + 0.007919159717857838, + 0.02162890136241913, + 0.006310255266726017, + 0.009073576889932156, + 0.010517860762774944, + 0.08945870399475098, + 0.010763514786958694, + 0.012116492725908756, + 0.009393158368766308, + 0.00947505235671997, + 0.009074890054762363, + 0.013804645277559757, + 0.009607478976249695, + 0.012406383641064167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.006591245997697115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181990098208189 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.0181990098208189, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181990098208189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10637042671442032, + 0.055803362280130386, + 0.08244118094444275, + 0.29384198784828186, + 0.07634324580430984, + 0.05077751353383064, + 0.031111573800444603, + 0.02273394726216793, + 0.23636628687381744, + 0.04421049728989601 + ], + "credit_map": [ + 0.012127022258937359, + 0.01246840413659811, + 0.007822278887033463, + 0.012055122293531895, + 0.015309575013816357, + 0.019372066482901573, + 0.011467601172626019, + 0.01574835553765297, + 0.007448611315339804, + 0.007439850829541683, + 0.007306491956114769, + 0.0063074566423892975, + 0.007573995273560286, + 0.006711732130497694, + 0.006915539037436247, + 0.006099681369960308, + 0.011736011132597923, + 0.012353413738310337, + 0.009009317494928837, + 0.009567770175635815, + 0.009092002175748348, + 0.010255354456603527, + 0.009701420553028584, + 0.010725893080234528, + 0.028797971084713936, + 0.033174920827150345, + 0.03947151079773903, + 0.02446543052792549, + 0.06420819461345673, + 0.03430758789181709, + 0.039732981473207474, + 0.029683399945497513, + 0.008337575010955334, + 0.00578280258923769, + 0.007718753069639206, + 0.01995195634663105, + 0.004595099948346615, + 0.008058138191699982, + 0.010808268561959267, + 0.011090652085840702, + 0.006897415034472942, + 0.00513699185103178, + 0.008244652301073074, + 0.006107523571699858, + 0.00548442080616951, + 0.005378620699048042, + 0.006150244735181332, + 0.0073776450008153915, + 0.004778637085109949, + 0.0039947605691850185, + 0.003176254453137517, + 0.00379621097818017, + 0.0038922422099858522, + 0.0035641691647469997, + 0.004621114581823349, + 0.0032881854567676783, + 0.003873182227835059, + 0.002874926198273897, + 0.0021567584481090307, + 0.0029323846101760864, + 0.002784338779747486, + 0.003282682504504919, + 0.0025948078837245703, + 0.0022348654456436634, + 0.09130532294511795, + 0.008458463475108147, + 0.007129728328436613, + 0.02165420912206173, + 0.0035226920153945684, + 0.008844819851219654, + 0.0072856019251048565, + 0.08816544711589813, + 0.00752358790487051, + 0.0039954050444066525, + 0.0028078213799744844, + 0.004131479188799858, + 0.00259232334792614, + 0.0060284994542598724, + 0.0035194656811654568, + 0.013611915521323681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0181990098208189 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370063751935959 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370063751935959, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370063751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15756720304489136, + 0.13654007017612457, + 0.0720793828368187, + 0.05528320372104645, + 0.10346664488315582, + 0.16313646733760834, + 0.02336106263101101, + 0.042176488786935806, + 0.10360243171453476, + 0.14278702437877655 + ], + "credit_map": [ + 0.020149709656834602, + 0.02204626426100731, + 0.013243093155324459, + 0.031604547053575516, + 0.018092896789312363, + 0.01743735559284687, + 0.016898013651371002, + 0.01809532195329666, + 0.018692532554268837, + 0.024738525971770287, + 0.020300185307860374, + 0.018021799623966217, + 0.014570474624633789, + 0.011348655447363853, + 0.014091022312641144, + 0.014776868745684624, + 0.011304520070552826, + 0.010248079895973206, + 0.007946545258164406, + 0.00591802503913641, + 0.007084496784955263, + 0.008805024437606335, + 0.008664424531161785, + 0.012108266353607178, + 0.005278094206005335, + 0.004777431488037109, + 0.008034221827983856, + 0.006816953420639038, + 0.00913082342594862, + 0.008401710540056229, + 0.007284710183739662, + 0.005559260956943035, + 0.010075520724058151, + 0.008032320998609066, + 0.010601307265460491, + 0.02396814152598381, + 0.007914187386631966, + 0.012366185896098614, + 0.014752711169421673, + 0.01575627736747265, + 0.020079122856259346, + 0.01618579775094986, + 0.025408916175365448, + 0.01889405958354473, + 0.019527260214090347, + 0.02118595875799656, + 0.019129740074276924, + 0.022725611925125122, + 0.003017494920641184, + 0.0032573796343058348, + 0.0025101385544985533, + 0.0029385881498456, + 0.002367987297475338, + 0.0031466020736843348, + 0.003424278926104307, + 0.002698594471439719, + 0.0054404051043093204, + 0.005972710903733969, + 0.0036996491253376007, + 0.006195993162691593, + 0.0053109354339540005, + 0.004678613040596247, + 0.006163795944303274, + 0.004714389331638813, + 0.03218185156583786, + 0.007098667323589325, + 0.005788121838122606, + 0.008876629173755646, + 0.005685142241418362, + 0.004925816785544157, + 0.0077330986969172955, + 0.031313106417655945, + 0.02266252040863037, + 0.015003427863121033, + 0.009037480689585209, + 0.01378630194813013, + 0.014593170955777168, + 0.01572669856250286, + 0.01342939492315054, + 0.03854803368449211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.03370063751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198780421167612 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198780421167612, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198780421167612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19581091403961182, + 0.059085663408041, + 0.08739646524190903, + 0.028660930693149567, + 0.03434496745467186, + 0.052889563143253326, + 0.014852993190288544, + 0.049494389444589615, + 0.2965969145298004, + 0.18086718022823334 + ], + "credit_map": [ + 0.026018910109996796, + 0.022440796718001366, + 0.014084212481975555, + 0.023662667721509933, + 0.026084646582603455, + 0.0286874920129776, + 0.019901303574442863, + 0.03493090346455574, + 0.007352775894105434, + 0.00816208403557539, + 0.007379837334156036, + 0.009308193810284138, + 0.007324007339775562, + 0.004692485556006432, + 0.008231747895479202, + 0.0066345324739813805, + 0.01068524457514286, + 0.013228658586740494, + 0.013503792695701122, + 0.009301647543907166, + 0.006643995642662048, + 0.008958224207162857, + 0.013162611983716488, + 0.011912294663488865, + 0.0033743204548954964, + 0.004285811446607113, + 0.0023998136166483164, + 0.0038647339679300785, + 0.004397079348564148, + 0.003298485651612282, + 0.0043104360811412334, + 0.0027302519883960485, + 0.0031170984730124474, + 0.00325034954585135, + 0.004197713918983936, + 0.0076678041368722916, + 0.0028102921787649393, + 0.0033518567215651274, + 0.005517754703760147, + 0.0044320980086922646, + 0.007630243897438049, + 0.004934205207973719, + 0.009606597013771534, + 0.006615061312913895, + 0.005721040535718203, + 0.005598928779363632, + 0.007010017056018114, + 0.005773474927991629, + 0.0025213584303855896, + 0.0014096334343776107, + 0.0015586231602355838, + 0.0013274572556838393, + 0.001737514277920127, + 0.0030672054272145033, + 0.0017022100510075688, + 0.0015289925504475832, + 0.005441879853606224, + 0.006285994779318571, + 0.004234835971146822, + 0.006444182246923447, + 0.006995351053774357, + 0.005973335355520248, + 0.007182077504694462, + 0.006936733610928059, + 0.10954315960407257, + 0.012056062929332256, + 0.01017578225582838, + 0.02553478255867958, + 0.007774264086037874, + 0.012206214480102062, + 0.010855277068912983, + 0.10845140367746353, + 0.025285275653004646, + 0.022775879129767418, + 0.009764861315488815, + 0.018139902502298355, + 0.0165395587682724, + 0.025509603321552277, + 0.01697816513478756, + 0.04587394371628761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.002198780421167612 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a massive Godzilla covered in rough, dark gray scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the ruins of a shattered city with skyscrapers lying beneath the subject's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Godzilla throw its head back with its jaws stretched wide while emitting a wave of energy from its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, atomic blue light pulsating from Godzilla's throat and mouth that illuminates the internal structure and casts a huge shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an overhead view that swivels in a 360-degree arc around the subject at breakneck speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31339287757873535 + ], + "opposite_logprob_tokens": [ + -1.3133928775787354 + ], + "logprob_reward": -0.31339287757873535, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31339287757873535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07945089042186737, + 0.05858968198299408, + 0.10259737074375153, + 0.03791573643684387, + 0.08497227728366852, + 0.06731070578098297, + 0.03690127655863762, + 0.033477313816547394, + 0.24769985675811768, + 0.2510848939418793 + ], + "credit_map": [ + 0.008628026582300663, + 0.008776417933404446, + 0.004648766480386257, + 0.008920404128730297, + 0.011195476166903973, + 0.013697929680347443, + 0.009115509688854218, + 0.014468356966972351, + 0.007743657100945711, + 0.009912303648889065, + 0.00895380973815918, + 0.007061763200908899, + 0.006420846562832594, + 0.004456155002117157, + 0.008004545234143734, + 0.006036605220288038, + 0.01215918455272913, + 0.01186398509889841, + 0.016495509073138237, + 0.010775777511298656, + 0.00850493274629116, + 0.011690936982631683, + 0.013907025568187237, + 0.017200032249093056, + 0.0036218753084540367, + 0.0031992709264159203, + 0.005080769304186106, + 0.006086396519094706, + 0.005222206469625235, + 0.005601557902991772, + 0.005496981553733349, + 0.003606675658375025, + 0.00945358257740736, + 0.006777211092412472, + 0.011818219907581806, + 0.024758506566286087, + 0.003977435640990734, + 0.005450881086289883, + 0.013941430486738682, + 0.008795006200671196, + 0.009136221371591091, + 0.0074906377121806145, + 0.009234508499503136, + 0.007862045429646969, + 0.008013383485376835, + 0.007372056134045124, + 0.008126722648739815, + 0.010075132362544537, + 0.0053095631301403046, + 0.005200116895139217, + 0.0030777952633798122, + 0.0036720538046211004, + 0.004735547583550215, + 0.005128065124154091, + 0.004960630089044571, + 0.004817505832761526, + 0.00483919121325016, + 0.005683060735464096, + 0.0034720622934401035, + 0.004605541937053204, + 0.004291853401809931, + 0.0036181912291795015, + 0.003804204985499382, + 0.0031632045283913612, + 0.0949561670422554, + 0.009438223205506802, + 0.007509955670684576, + 0.02226993255317211, + 0.004522115457803011, + 0.009263386949896812, + 0.007775544188916683, + 0.09196453541517258, + 0.0449959896504879, + 0.023277364671230316, + 0.01362884882837534, + 0.02418169379234314, + 0.01657368242740631, + 0.03098933957517147, + 0.01956324651837349, + 0.07787472754716873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.31339287757873535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_tensor([9643, 9643])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_tensor([9643, 9643])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f992b04cb8e83083d365d0dd1a81618755873140 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_tensor([9643, 9643])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot focuses with macroscopic precision on a translucent icicle suspended from the weathered, dark wooden ledge of an exterior window frame. The low, golden light of the winter afternoon sun casts a shadow on the icicle, revealing its crystalline structure and the fine frost on its surface. At its sharp tip, a perfectly spherical droplet of meltwater deliberately expands, its surface tension creating a clear, inverted, blurred snowscape in the background. Under its own weight, the droplet stretches downward, its shape elongating from a sphere to a teardrop. The thin thread connecting it to the icicle shimmers for a moment before inevitably breaking, and the droplet quickly and silently falls from the bottom of the frame, temporarily dampening the tip of the icicle. The scene is imbued with a tranquil, quiet atmosphere, capturing the subtle transitional moment with realistic detail.", + "seed": 24333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_9643_seed_24333_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04654169925488531, + "vlm_logprob": -0.04654169925488531, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.04654169925488531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.057180196046829224 + ], + "opposite_logprob_tokens": [ + -4.807180404663086 + ], + "logprob_reward": -0.057180196046829224, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.057180196046829224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08171660453081131, + 0.024549312889575958, + 0.09625935554504395, + 0.06525757908821106, + 0.04662924259901047, + 0.06381848454475403, + 0.09846164286136627, + 0.33449363708496094, + 0.12480935454368591, + 0.06400474905967712 + ], + "credit_map": [ + 0.011790095828473568, + 0.010259754955768585, + 0.015597132965922356, + 0.007896298542618752, + 0.007739384192973375, + 0.0072729396633803844, + 0.006057227961719036, + 0.015103779733181, + 0.0037984682712703943, + 0.003322526579722762, + 0.003540806472301483, + 0.002678626449778676, + 0.002626552479341626, + 0.003068386809900403, + 0.0019651432521641254, + 0.0035488037392497063, + 0.012560681439936161, + 0.009470377117395401, + 0.02536015398800373, + 0.0059351129457354546, + 0.003674247534945607, + 0.0098265977576375, + 0.0035660509020090103, + 0.025866134092211723, + 0.008921438828110695, + 0.005932159256190062, + 0.018055006861686707, + 0.004140895325690508, + 0.002702397760003805, + 0.005393247120082378, + 0.003554136026650667, + 0.01655830256640911, + 0.004785718396306038, + 0.005766858346760273, + 0.007399069145321846, + 0.004740051459521055, + 0.0076227313838899136, + 0.004020409658551216, + 0.005374141037464142, + 0.006920265965163708, + 0.00645556440576911, + 0.007007124368101358, + 0.010640442371368408, + 0.006648819427937269, + 0.008492600172758102, + 0.00709003210067749, + 0.005710349418222904, + 0.011773561127483845, + 0.014263798482716084, + 0.02049858309328556, + 0.011897067539393902, + 0.010457487776875496, + 0.011566617526113987, + 0.008835825137794018, + 0.009250792674720287, + 0.011691469699144363, + 0.023090306669473648, + 0.06221611797809601, + 0.04303423687815666, + 0.04140999913215637, + 0.03326119855046272, + 0.048546746373176575, + 0.04359929636120796, + 0.03933578357100487, + 0.008442830294370651, + 0.018072010949254036, + 0.013068034313619137, + 0.015920571982860565, + 0.0142294866964221, + 0.014473785646259785, + 0.025142526254057884, + 0.015460115857422352, + 0.005243634805083275, + 0.006266403943300247, + 0.007594125811010599, + 0.015245056711137295, + 0.005069791339337826, + 0.004791300278156996, + 0.007535269483923912, + 0.012259170413017273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.057180196046829224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001959072658792138 + ], + "opposite_logprob_tokens": [ + -6.251958847045898 + ], + "logprob_reward": -0.001959072658792138, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001959072658792138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20560812950134277, + 0.1076785996556282, + 0.05073459818959236, + 0.07143200933933258, + 0.11676541715860367, + 0.12635919451713562, + 0.08533108979463577, + 0.050498683005571365, + 0.0745406523346901, + 0.11105157434940338 + ], + "credit_map": [ + 0.020688461139798164, + 0.014679656364023685, + 0.057272329926490784, + 0.013980647549033165, + 0.018933095037937164, + 0.01877009868621826, + 0.01300812792032957, + 0.04827570170164108, + 0.006829152815043926, + 0.00633806549012661, + 0.03306375816464424, + 0.006578666158020496, + 0.005632973741739988, + 0.013140288181602955, + 0.00557349156588316, + 0.03052220493555069, + 0.005271164234727621, + 0.00551063846796751, + 0.010984556749463081, + 0.004107357002794743, + 0.004329838789999485, + 0.006430049426853657, + 0.0036766217090189457, + 0.010424376465380192, + 0.007447039242833853, + 0.007672002073377371, + 0.016927670687437057, + 0.006485924124717712, + 0.005394714418798685, + 0.0056990813463926315, + 0.007154832128435373, + 0.014650749042630196, + 0.022721795365214348, + 0.022492634132504463, + 0.015418151393532753, + 0.010967028327286243, + 0.009412371553480625, + 0.009947968646883965, + 0.007181044667959213, + 0.018624423071742058, + 0.018310442566871643, + 0.016812104731798172, + 0.02724466472864151, + 0.010159793309867382, + 0.008946208283305168, + 0.00991048477590084, + 0.0069814277812838554, + 0.027994075790047646, + 0.008304106071591377, + 0.01750466786324978, + 0.017716240137815475, + 0.009083343669772148, + 0.0077132913284003735, + 0.007427975069731474, + 0.005894117522984743, + 0.011687343940138817, + 0.0046296874061226845, + 0.007409092504531145, + 0.007103705313056707, + 0.007196935825049877, + 0.005759949330240488, + 0.004291726276278496, + 0.007815900258719921, + 0.006291686091572046, + 0.005011168774217367, + 0.009601124562323093, + 0.007803009822964668, + 0.010832727886736393, + 0.0062258574180305, + 0.007634022738784552, + 0.016845138743519783, + 0.010587603785097599, + 0.009486839175224304, + 0.011993455700576305, + 0.012796442955732346, + 0.021810149773955345, + 0.016490934416651726, + 0.012615051120519638, + 0.011291047558188438, + 0.014567648991942406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.001959072658792138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819748803973198 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.01819748803973198, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819748803973198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11002540588378906, + 0.09512077271938324, + 0.04351985082030296, + 0.14299698173999786, + 0.14512965083122253, + 0.11953403800725937, + 0.05440784990787506, + 0.0897420197725296, + 0.07760785520076752, + 0.1219155415892601 + ], + "credit_map": [ + 0.007716525346040726, + 0.009707138873636723, + 0.029467353597283363, + 0.007317556533962488, + 0.011798384599387646, + 0.011834364384412766, + 0.007780588697642088, + 0.024403495714068413, + 0.00681620929390192, + 0.004947538021951914, + 0.03139050677418709, + 0.006005721166729927, + 0.006880395580083132, + 0.004989370703697205, + 0.006151142995804548, + 0.02793988771736622, + 0.0038906987756490707, + 0.005441388580948114, + 0.005465545225888491, + 0.004344997461885214, + 0.005525995511561632, + 0.007413956802338362, + 0.005850410088896751, + 0.005586858838796616, + 0.01118694618344307, + 0.010495842434465885, + 0.030965713784098625, + 0.013928640633821487, + 0.010560574941337109, + 0.02421443536877632, + 0.012140197679400444, + 0.02950463257730007, + 0.018336251378059387, + 0.01572536490857601, + 0.019357120618224144, + 0.014085332863032818, + 0.018335741013288498, + 0.015234400518238544, + 0.02143380604684353, + 0.02262163907289505, + 0.015193521976470947, + 0.010887597687542439, + 0.030357621610164642, + 0.009036296978592873, + 0.008296315558254719, + 0.008418906480073929, + 0.007019997574388981, + 0.03032378852367401, + 0.006826450116932392, + 0.01219691801816225, + 0.0063279238529503345, + 0.006006489973515272, + 0.005918697454035282, + 0.004489766899496317, + 0.005867069587111473, + 0.006774535868316889, + 0.0050720577128231525, + 0.01749388687312603, + 0.010617358610033989, + 0.009194286540150642, + 0.008557643741369247, + 0.02172279730439186, + 0.007815160788595676, + 0.009268827736377716, + 0.005741429049521685, + 0.007234473247081041, + 0.009808476082980633, + 0.012033827602863312, + 0.010019022971391678, + 0.005949436686933041, + 0.016955768689513206, + 0.00986542273312807, + 0.008281858637928963, + 0.015199162065982819, + 0.015087539330124855, + 0.021345432847738266, + 0.02256009355187416, + 0.012870092876255512, + 0.010382609441876411, + 0.016188757494091988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.01819748803973198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10004035383462906, + 0.1075674295425415, + 0.021719157695770264, + 0.05311828851699829, + 0.11036957055330276, + 0.1362588256597519, + 0.07638949155807495, + 0.19229505956172943, + 0.14860041439533234, + 0.05364147573709488 + ], + "credit_map": [ + 0.01072755642235279, + 0.00815910380333662, + 0.024125924333930016, + 0.005868677049875259, + 0.00586268724873662, + 0.014343386515974998, + 0.004682452417910099, + 0.026270557194948196, + 0.007640364114195108, + 0.006128068547695875, + 0.03303319215774536, + 0.007162779103964567, + 0.008619273081421852, + 0.007864272221922874, + 0.0074526299722492695, + 0.029666846618056297, + 0.0038125908467918634, + 0.003243393497541547, + 0.003334491280838847, + 0.0023200539872050285, + 0.0025421043392270803, + 0.0021723932586610317, + 0.001314929104410112, + 0.0029792014975100756, + 0.0069146729074418545, + 0.005389302968978882, + 0.009810873307287693, + 0.004803611431270838, + 0.005837824195623398, + 0.0062324171885848045, + 0.004009468015283346, + 0.010120113380253315, + 0.015664175152778625, + 0.013056999072432518, + 0.01655840501189232, + 0.013553154654800892, + 0.012816826812922955, + 0.00995535310357809, + 0.009888614527881145, + 0.01887604594230652, + 0.014822431840002537, + 0.013977080583572388, + 0.030332013964653015, + 0.013611186295747757, + 0.012409987859427929, + 0.009951653890311718, + 0.01035735197365284, + 0.030797110870480537, + 0.009657717309892178, + 0.01603364385664463, + 0.008517774753272533, + 0.008144273422658443, + 0.009712133556604385, + 0.006843386217951775, + 0.00872762966901064, + 0.008752930909395218, + 0.011582836508750916, + 0.03572937846183777, + 0.027383053675293922, + 0.02098466269671917, + 0.018126944079995155, + 0.03894960135221481, + 0.01689610444009304, + 0.022642482072114944, + 0.008677207864820957, + 0.019767876714468002, + 0.018637726083397865, + 0.03446926549077034, + 0.011650614440441132, + 0.01672341302037239, + 0.020234616473317146, + 0.018439697101712227, + 0.005003819242119789, + 0.00708241481333971, + 0.0070174578577280045, + 0.009367766790091991, + 0.007519910577684641, + 0.005877988878637552, + 0.005598823074251413, + 0.00617329403758049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282649755477905 + ], + "opposite_logprob_tokens": [ + -2.017826557159424 + ], + "logprob_reward": -0.14282649755477905, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282649755477905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12497719377279282, + 0.04013243690133095, + 0.059042613953351974, + 0.054872576147317886, + 0.07521723210811615, + 0.08349254727363586, + 0.14939828217029572, + 0.17311552166938782, + 0.19635958969593048, + 0.04339196905493736 + ], + "credit_map": [ + 0.019142506644129753, + 0.017812279984354973, + 0.024004895240068436, + 0.010126486420631409, + 0.00991497840732336, + 0.013030393980443478, + 0.007137920707464218, + 0.023807736113667488, + 0.005183002911508083, + 0.004358120262622833, + 0.007359674666076899, + 0.002405531005933881, + 0.0038078119978308678, + 0.0065175616182386875, + 0.0034263257402926683, + 0.007074414286762476, + 0.007327216677367687, + 0.005979252513498068, + 0.013371254317462444, + 0.003907712642103434, + 0.0038832654245197773, + 0.007778157480061054, + 0.003771689487621188, + 0.013024070300161839, + 0.009232399053871632, + 0.008123018778860569, + 0.010941521264612675, + 0.00378523045219481, + 0.004597906954586506, + 0.0042701056227087975, + 0.0031136157922446728, + 0.010808780789375305, + 0.006854221224784851, + 0.00697710458189249, + 0.01486123912036419, + 0.008340860716998577, + 0.007677322253584862, + 0.008164037019014359, + 0.007854861207306385, + 0.014487586915493011, + 0.01262641977518797, + 0.011292577721178532, + 0.01142881065607071, + 0.009805123321712017, + 0.008798714727163315, + 0.009230002760887146, + 0.007860247045755386, + 0.01245064940303564, + 0.02275707572698593, + 0.04125062748789787, + 0.016752297058701515, + 0.014740723185241222, + 0.013371942564845085, + 0.013745916076004505, + 0.010476377792656422, + 0.016303330659866333, + 0.010190809145569801, + 0.025067294016480446, + 0.021551121026277542, + 0.01728226989507675, + 0.019906960427761078, + 0.031830910593271255, + 0.02560686506330967, + 0.021679319441318512, + 0.012972289696335793, + 0.030476123094558716, + 0.018700189888477325, + 0.035251814872026443, + 0.01733863726258278, + 0.02448432706296444, + 0.025879597291350365, + 0.03125661239027977, + 0.0038365316577255726, + 0.005497719161212444, + 0.004550803918391466, + 0.006758112460374832, + 0.006448142230510712, + 0.006052702199667692, + 0.004325140733271837, + 0.0059228213503956795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.14282649755477905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot focuses with macroscopic precision on a translucent icicle suspended from the weathered, dark wooden ledge of an exterior window frame. The low, golden light of the winter afternoon sun casts a shadow on the icicle, revealing its crystalline structure and the fine frost on its surface. At its sharp tip, a perfectly spherical droplet of meltwater deliberately expands, its surface tension creating a clear, inverted, blurred snowscape in the background. Under its own weight, the droplet stretches downward, its shape elongating from a sphere to a teardrop. The thin thread connecting it to the icicle shimmers for a moment before inevitably breaking, and the droplet quickly and silently falls from the bottom of the frame, temporarily dampening the tip of the icicle. The scene is imbued with a tranquil, quiet atmosphere, capturing the subtle transitional moment with realistic detail.", + "seed": 24334, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_9643_seed_24334_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2106099876575172, + "vlm_logprob": -0.2106099876575172, + "vlm_yes_no_margin": 3.6249999582767485, + "vlm_token_credit": -0.2106099876575172, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04226892814040184 + ], + "opposite_logprob_tokens": [ + -4.417268753051758 + ], + "logprob_reward": -0.04226892814040184, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.04226892814040184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044010259211063385, + 0.05727435648441315, + 0.047899775207042694, + 0.040014270693063736, + 0.06037573143839836, + 0.1582048535346985, + 0.1554841548204422, + 0.32313910126686096, + 0.08295905590057373, + 0.030638443306088448 + ], + "credit_map": [ + 0.0047732265666127205, + 0.004011708777397871, + 0.006116955075412989, + 0.005037689581513405, + 0.0048163882456719875, + 0.005878196097910404, + 0.008249621838331223, + 0.005126475356519222, + 0.005848311819136143, + 0.006142855621874332, + 0.004578078631311655, + 0.005093352869153023, + 0.013969090767204762, + 0.007214638404548168, + 0.006376056000590324, + 0.008051975630223751, + 0.005566313862800598, + 0.0065353973768651485, + 0.005602989811450243, + 0.0074673006311059, + 0.005651311483234167, + 0.005254318006336689, + 0.00571753503754735, + 0.0061046043410897255, + 0.0034684683196246624, + 0.0055350870825350285, + 0.005151732824742794, + 0.005557411350309849, + 0.006581563502550125, + 0.004171955399215221, + 0.004615554586052895, + 0.00493249949067831, + 0.005509157199412584, + 0.00829757284373045, + 0.007119316607713699, + 0.01324046216905117, + 0.007151456084102392, + 0.005045257043093443, + 0.007101764436811209, + 0.00691074226051569, + 0.016532741487026215, + 0.027022680267691612, + 0.016029970720410347, + 0.021774906665086746, + 0.02238926663994789, + 0.011111730709671974, + 0.019439388066530228, + 0.023904168978333473, + 0.01392881479114294, + 0.020319001749157906, + 0.015491122379899025, + 0.024372171610593796, + 0.019381005316972733, + 0.011044343002140522, + 0.02099374495446682, + 0.029953939840197563, + 0.03132238984107971, + 0.08400560170412064, + 0.018563734367489815, + 0.022977545857429504, + 0.06590026617050171, + 0.02053292840719223, + 0.02429240569472313, + 0.05554423853754997, + 0.007149945013225079, + 0.015561478212475777, + 0.011144179850816727, + 0.006605683360248804, + 0.014297530986368656, + 0.007870177738368511, + 0.008890378288924694, + 0.011439688503742218, + 0.003370506688952446, + 0.0039753299206495285, + 0.0038972923066467047, + 0.0023890931624919176, + 0.003991857171058655, + 0.002983186161145568, + 0.004589652642607689, + 0.005441523157060146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.04226892814040184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002216027583926916 + ], + "opposite_logprob_tokens": [ + -6.12721586227417 + ], + "logprob_reward": -0.002216027583926916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002216027583926916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10685202479362488, + 0.20497390627861023, + 0.05927170813083649, + 0.04621661454439163, + 0.03551110625267029, + 0.06545615196228027, + 0.0774165466427803, + 0.16835817694664001, + 0.13487295806407928, + 0.10107075423002243 + ], + "credit_map": [ + 0.008233869448304176, + 0.017805293202400208, + 0.012787111103534698, + 0.014200485311448574, + 0.017970135435461998, + 0.0068894969299435616, + 0.011958755552768707, + 0.017006896436214447, + 0.01904197596013546, + 0.025537146255373955, + 0.01628888212144375, + 0.017423290759325027, + 0.05571218580007553, + 0.022358611226081848, + 0.01872655190527439, + 0.02988528087735176, + 0.004577601328492165, + 0.007965781725943089, + 0.0088192168623209, + 0.00911531038582325, + 0.008797701448202133, + 0.006027700379490852, + 0.006560636218637228, + 0.007407762575894594, + 0.004061080515384674, + 0.00560305779799819, + 0.005864952225238085, + 0.00878102332353592, + 0.007331398781388998, + 0.003974503837525845, + 0.004893700126558542, + 0.005706902127712965, + 0.0033344165422022343, + 0.005735377315431833, + 0.0038281793240457773, + 0.004498328547924757, + 0.005693041253834963, + 0.004334581084549427, + 0.0034598647616803646, + 0.0046273209154605865, + 0.007047462742775679, + 0.012516842223703861, + 0.008387728594243526, + 0.006942412815988064, + 0.008741048164665699, + 0.006532183848321438, + 0.008447270840406418, + 0.006841206457465887, + 0.0053352597169578075, + 0.01333948876708746, + 0.00826929323375225, + 0.012914823368191719, + 0.010763748548924923, + 0.008240221999585629, + 0.007972915656864643, + 0.010580796748399734, + 0.014399905689060688, + 0.04097646474838257, + 0.012596176005899906, + 0.011488102376461029, + 0.031716927886009216, + 0.01468762569129467, + 0.014753947965800762, + 0.02773904614150524, + 0.012520957738161087, + 0.02318187989294529, + 0.014921548776328564, + 0.01105333399027586, + 0.021866347640752792, + 0.012000206857919693, + 0.014791239984333515, + 0.024537449702620506, + 0.010755718685686588, + 0.014410015195608139, + 0.010088123381137848, + 0.007383404299616814, + 0.022108951583504677, + 0.010697375051677227, + 0.012302401475608349, + 0.013324773870408535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.002216027583926916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608244515955448 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608244515955448, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608244515955448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03964800760149956, + 0.05563736706972122, + 0.051865775138139725, + 0.031042087823152542, + 0.06523182988166809, + 0.05082380771636963, + 0.05638003349304199, + 0.5966441631317139, + 0.028514768928289413, + 0.024212170392274857 + ], + "credit_map": [ + 0.004819901660084724, + 0.004582579247653484, + 0.0053809876553714275, + 0.005243184510618448, + 0.004584835842251778, + 0.004398653283715248, + 0.005642763804644346, + 0.004995101597160101, + 0.005838642828166485, + 0.005919264629483223, + 0.0041917297057807446, + 0.004048182629048824, + 0.014661401510238647, + 0.0067621986381709576, + 0.005931254010647535, + 0.008284692652523518, + 0.005384722258895636, + 0.00786511693149805, + 0.005974060855805874, + 0.009005913510918617, + 0.006782022304832935, + 0.004147638566792011, + 0.006204993464052677, + 0.0065013044513762, + 0.0029685802292078733, + 0.004087324719876051, + 0.004159315023571253, + 0.0051632532849907875, + 0.0037335131783038378, + 0.0031795704271644354, + 0.0031991044525057077, + 0.0045514279045164585, + 0.005229703616350889, + 0.010268385522067547, + 0.008432211354374886, + 0.009854687377810478, + 0.00996619276702404, + 0.006269233301281929, + 0.007594205904752016, + 0.007617210503667593, + 0.004653157200664282, + 0.009647532366216183, + 0.0058866036124527454, + 0.006574390456080437, + 0.007548661902546883, + 0.00478637870401144, + 0.005514644552022219, + 0.006212439853698015, + 0.005052220541983843, + 0.00633119186386466, + 0.007938372902572155, + 0.01125673484057188, + 0.005729760508984327, + 0.003261210396885872, + 0.006623418070375919, + 0.010187123902142048, + 0.050939712673425674, + 0.1559104472398758, + 0.03808233514428139, + 0.04155783727765083, + 0.11308395862579346, + 0.04261810705065727, + 0.049938518553972244, + 0.1045132502913475, + 0.002653767354786396, + 0.004128426779061556, + 0.0034673381596803665, + 0.0022272150963544846, + 0.0035605947487056255, + 0.0022653834894299507, + 0.00489591620862484, + 0.005316126625984907, + 0.0019918298348784447, + 0.003356021596118808, + 0.002611069241538644, + 0.0017391907749697566, + 0.004059532657265663, + 0.0028433515690267086, + 0.00313859386369586, + 0.004472580272704363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.01608244515955448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188709393143654 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188709393143654, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188709393143654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038207728415727615, + 0.09006620198488235, + 0.02811945602297783, + 0.04635370150208473, + 0.06385745108127594, + 0.20924760401248932, + 0.2530173063278198, + 0.15538576245307922, + 0.0635317713022232, + 0.05221296474337578 + ], + "credit_map": [ + 0.004316587932407856, + 0.004116963595151901, + 0.004787903744727373, + 0.004822960589081049, + 0.006084672175347805, + 0.003948527853935957, + 0.0049923136830329895, + 0.005137798376381397, + 0.009241237305104733, + 0.010105492547154427, + 0.007686343044042587, + 0.007169339805841446, + 0.02407095395028591, + 0.0101768234744668, + 0.008703555911779404, + 0.012912456877529621, + 0.002846094546839595, + 0.003897876711562276, + 0.0032519286032766104, + 0.0033214495051652193, + 0.005812382325530052, + 0.0029142480343580246, + 0.002120563993230462, + 0.003954913001507521, + 0.0040999785996973515, + 0.007040089461952448, + 0.004929931368678808, + 0.007380012422800064, + 0.005941132083535194, + 0.004463803023099899, + 0.005517304874956608, + 0.006981447339057922, + 0.006063870154321194, + 0.008093520998954773, + 0.008398860692977905, + 0.009651806205511093, + 0.00978042371571064, + 0.007147102616727352, + 0.007987639866769314, + 0.006734220311045647, + 0.020527170971035957, + 0.03130006417632103, + 0.023145632818341255, + 0.033247023820877075, + 0.027332128956913948, + 0.013836324214935303, + 0.02601696364581585, + 0.0338422991335392, + 0.027519727125763893, + 0.04754339158535004, + 0.018940171226859093, + 0.03149638697504997, + 0.040526386350393295, + 0.02232513576745987, + 0.025951595976948738, + 0.03871453180909157, + 0.013482101261615753, + 0.039433833211660385, + 0.008582668378949165, + 0.009674345143139362, + 0.03473953902721405, + 0.014748421497642994, + 0.012410554103553295, + 0.02231430821120739, + 0.006393491756170988, + 0.005882605444639921, + 0.008141650818288326, + 0.005066342186182737, + 0.007672000210732222, + 0.008541754446923733, + 0.008807890117168427, + 0.013026034459471703, + 0.004947121720761061, + 0.005440704990178347, + 0.005867142230272293, + 0.0037808059714734554, + 0.006581016816198826, + 0.0051049417816102505, + 0.007186422124505043, + 0.013304810039699078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.018188709393143654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742938280105591 + ], + "opposite_logprob_tokens": [ + -0.47429385781288147 + ], + "logprob_reward": -0.9742938280105591, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742938280105591, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15096303820610046, + 0.06097745522856712, + 0.028920866549015045, + 0.060267943888902664, + 0.04755902290344238, + 0.12638422846794128, + 0.17133231461048126, + 0.1964757889509201, + 0.10515627264976501, + 0.05196302384138107 + ], + "credit_map": [ + 0.011579830199480057, + 0.01828177645802498, + 0.016052305698394775, + 0.018524810671806335, + 0.024833394214510918, + 0.01812341809272766, + 0.02400972880423069, + 0.019557766616344452, + 0.005258842371404171, + 0.004931189119815826, + 0.005175817757844925, + 0.007170023862272501, + 0.016405263915657997, + 0.0068592471070587635, + 0.006180995609611273, + 0.008996075950562954, + 0.003397905733436346, + 0.0031879968009889126, + 0.0035643649753183126, + 0.003892558393999934, + 0.003608028870075941, + 0.004127716179937124, + 0.003005777485668659, + 0.004136519506573677, + 0.005712626501917839, + 0.008235092274844646, + 0.007719953544437885, + 0.007024222053587437, + 0.010954305529594421, + 0.007074319291859865, + 0.0064286841079592705, + 0.0071187405847013, + 0.003305165097117424, + 0.007636289112269878, + 0.0060269637033343315, + 0.008138992823660374, + 0.007172410376369953, + 0.0049756537191569805, + 0.004891682416200638, + 0.005411868449300528, + 0.013449054211378098, + 0.021662188693881035, + 0.015906844288110733, + 0.015449910424649715, + 0.019116081297397614, + 0.01205875352025032, + 0.014342754147946835, + 0.014398644678294659, + 0.01332217175513506, + 0.028989851474761963, + 0.014384151436388493, + 0.02766409143805504, + 0.023384595289826393, + 0.013562612235546112, + 0.028519466519355774, + 0.02150537259876728, + 0.017796754837036133, + 0.03988206759095192, + 0.014099310152232647, + 0.014738121069967747, + 0.049236565828323364, + 0.017154239118099213, + 0.01411587093025446, + 0.029452849179506302, + 0.010309870354831219, + 0.015236379578709602, + 0.012976452708244324, + 0.009127633646130562, + 0.013738254085183144, + 0.011118294671177864, + 0.008051336742937565, + 0.024598056450486183, + 0.004204734694212675, + 0.008568095043301582, + 0.004053326323628426, + 0.004100440535694361, + 0.010445022955536842, + 0.005818454548716545, + 0.006239349953830242, + 0.008533603511750698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.9742938280105591 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, static shot focuses with macroscopic precision on a translucent icicle suspended from the weathered, dark wooden ledge of an exterior window frame. The low, golden light of the winter afternoon sun casts a shadow on the icicle, revealing its crystalline structure and the fine frost on its surface. At its sharp tip, a perfectly spherical droplet of meltwater deliberately expands, its surface tension creating a clear, inverted, blurred snowscape in the background. Under its own weight, the droplet stretches downward, its shape elongating from a sphere to a teardrop. The thin thread connecting it to the icicle shimmers for a moment before inevitably breaking, and the droplet quickly and silently falls from the bottom of the frame, temporarily dampening the tip of the icicle. The scene is imbued with a tranquil, quiet atmosphere, capturing the subtle transitional moment with realistic detail.", + "seed": 24335, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_9643_seed_24335_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23314054822549224, + "vlm_logprob": -0.23314054822549224, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.23314054822549224, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07286958396434784 + ], + "opposite_logprob_tokens": [ + -4.572869777679443 + ], + "logprob_reward": -0.07286958396434784, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.07286958396434784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1282978057861328, + 0.07427269965410233, + 0.11804470419883728, + 0.0756920799612999, + 0.20034264028072357, + 0.06405400484800339, + 0.08699749410152435, + 0.07544919103384018, + 0.06734172999858856, + 0.10950767248868942 + ], + "credit_map": [ + 0.008740153163671494, + 0.01184821967035532, + 0.01599208265542984, + 0.012048780918121338, + 0.014611390419304371, + 0.015051265247166157, + 0.02922632172703743, + 0.02077958732843399, + 0.0068910508416593075, + 0.006285491399466991, + 0.008240466006100178, + 0.007158576976507902, + 0.006595791783183813, + 0.010087748058140278, + 0.01784263364970684, + 0.011170940473675728, + 0.0074367341585457325, + 0.005235367454588413, + 0.01158283930271864, + 0.013081253506243229, + 0.006825346499681473, + 0.006918048951774836, + 0.043352216482162476, + 0.023612890392541885, + 0.007123391143977642, + 0.008440216071903706, + 0.012074037455022335, + 0.012155880220234394, + 0.0049810814671218395, + 0.009750529192388058, + 0.010234951041638851, + 0.010931992903351784, + 0.015162333846092224, + 0.013137737289071083, + 0.014765871688723564, + 0.015005400404334068, + 0.03931070491671562, + 0.023741846904158592, + 0.0461801141500473, + 0.03303862363100052, + 0.004793388303369284, + 0.00474288547411561, + 0.007602738682180643, + 0.006790810730308294, + 0.008182598277926445, + 0.008113976567983627, + 0.013518628664314747, + 0.010308975353837013, + 0.00484848627820611, + 0.0046981205232441425, + 0.006624252535402775, + 0.0069082360714674, + 0.01041510608047247, + 0.006395634729415178, + 0.030178222805261612, + 0.01692943647503853, + 0.00764009915292263, + 0.004258668515831232, + 0.016721021384000778, + 0.010934388265013695, + 0.006721081677824259, + 0.0080869160592556, + 0.009118184447288513, + 0.011968830600380898, + 0.004920552019029856, + 0.006034931167960167, + 0.007795849349349737, + 0.005412096623331308, + 0.008583832532167435, + 0.009109752252697945, + 0.013289171271026134, + 0.01219555176794529, + 0.010930013842880726, + 0.012487014755606651, + 0.011623070575296879, + 0.009237583726644516, + 0.014693429693579674, + 0.012339148670434952, + 0.01876887120306492, + 0.019428543746471405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07286958396434784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002220309805124998 + ], + "opposite_logprob_tokens": [ + -6.127220153808594 + ], + "logprob_reward": -0.002220309805124998, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002220309805124998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.156072199344635, + 0.1195167526602745, + 0.08026448637247086, + 0.08300212025642395, + 0.05616438016295433, + 0.06788718700408936, + 0.05609662085771561, + 0.11467952281236649, + 0.062267452478408813, + 0.2040492445230484 + ], + "credit_map": [ + 0.013070446439087391, + 0.01943839341402054, + 0.01649087481200695, + 0.016092805191874504, + 0.01626313477754593, + 0.029414238408207893, + 0.021212194114923477, + 0.0240901131182909, + 0.010989743284881115, + 0.009861255064606667, + 0.015331830829381943, + 0.011148769408464432, + 0.009895442984998226, + 0.018096638843417168, + 0.023770606145262718, + 0.020422469824552536, + 0.006277304142713547, + 0.004094171803444624, + 0.012629259377717972, + 0.010329099372029305, + 0.004893862642347813, + 0.007935694418847561, + 0.01954060234129429, + 0.014564494602382183, + 0.007320022210478783, + 0.008948705159127712, + 0.014636672101914883, + 0.013047531247138977, + 0.00529283145442605, + 0.011015579104423523, + 0.008537935093045235, + 0.014202846214175224, + 0.0046695563942193985, + 0.00293135829269886, + 0.007220390718430281, + 0.007121808361262083, + 0.005685493815690279, + 0.006420188117772341, + 0.012930330820381641, + 0.009185257367789745, + 0.004813154693692923, + 0.0049408297054469585, + 0.006116958335042, + 0.005264758598059416, + 0.011793632991611958, + 0.005537542514503002, + 0.016870474442839622, + 0.012549838982522488, + 0.005210707895457745, + 0.004267633426934481, + 0.010071590542793274, + 0.006930181756615639, + 0.004977019503712654, + 0.003981625661253929, + 0.011944903992116451, + 0.008712957613170147, + 0.008959654718637466, + 0.007718134671449661, + 0.011044892482459545, + 0.008088372647762299, + 0.020149804651737213, + 0.011390578001737595, + 0.024920564144849777, + 0.022407522425055504, + 0.005197192542254925, + 0.00666767405346036, + 0.00555718457326293, + 0.005790441296994686, + 0.008804920129477978, + 0.009249716065824032, + 0.011021727696061134, + 0.00997859612107277, + 0.012520506978034973, + 0.017938287928700447, + 0.018089652061462402, + 0.012547685764729977, + 0.019819434732198715, + 0.016132308170199394, + 0.05360223725438118, + 0.05339914560317993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.002220309805124998 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601926371455193 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020601926371455193, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601926371455193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2589271366596222, + 0.10598412156105042, + 0.05740624666213989, + 0.06114788353443146, + 0.07154573500156403, + 0.06631769984960556, + 0.03322925418615341, + 0.09046059846878052, + 0.05448908731341362, + 0.2004922479391098 + ], + "credit_map": [ + 0.01760336570441723, + 0.023816142231225967, + 0.022516630589962006, + 0.0237251129001379, + 0.02569236047565937, + 0.033059876412153244, + 0.06751061975955963, + 0.04500303417444229, + 0.006337183527648449, + 0.008984346874058247, + 0.007282375358045101, + 0.009494615718722343, + 0.018789686262607574, + 0.008728569373488426, + 0.027493523433804512, + 0.018873803317546844, + 0.004493982531130314, + 0.003378241090103984, + 0.006736500654369593, + 0.005686040967702866, + 0.0036468703765422106, + 0.00438280263915658, + 0.01880587823688984, + 0.010275923646986485, + 0.005496706813573837, + 0.004517547786235809, + 0.009964792989194393, + 0.007015386130660772, + 0.005046879407018423, + 0.007266059517860413, + 0.011075795628130436, + 0.010764708742499352, + 0.008121808990836143, + 0.005227180663496256, + 0.006505994591861963, + 0.004914590623229742, + 0.0079275444149971, + 0.010040849447250366, + 0.017310254275798798, + 0.01149750966578722, + 0.004649144131690264, + 0.00583110423758626, + 0.00675647146999836, + 0.005470579955726862, + 0.013835483230650425, + 0.00762708205729723, + 0.011385980062186718, + 0.010761846788227558, + 0.0034992166329175234, + 0.0020962487906217575, + 0.0034028547815978527, + 0.0031604385003447533, + 0.0036305144894868135, + 0.0033620260655879974, + 0.008418908342719078, + 0.005659046582877636, + 0.007861562073230743, + 0.005641161929816008, + 0.015414465218782425, + 0.008980616927146912, + 0.011343525722622871, + 0.012592361308634281, + 0.014141229912638664, + 0.014485659077763557, + 0.003923668526113033, + 0.007493994664400816, + 0.006023531313985586, + 0.0048696077428758144, + 0.006093982607126236, + 0.006791912484914064, + 0.010439183562994003, + 0.008853199891746044, + 0.010999657213687897, + 0.011216270737349987, + 0.020621273666620255, + 0.01152883842587471, + 0.015011638402938843, + 0.012809663079679012, + 0.05938821658492088, + 0.05891668424010277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.020601926371455193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607622765004635 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.01607622765004635, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607622765004635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07842046767473221, + 0.11792504787445068, + 0.05151475965976715, + 0.03085262142121792, + 0.09275707602500916, + 0.0956413671374321, + 0.12503862380981445, + 0.231352299451828, + 0.12068291008472443, + 0.05581482872366905 + ], + "credit_map": [ + 0.005772328935563564, + 0.010766733437776566, + 0.008906311355531216, + 0.00849137268960476, + 0.008426475338637829, + 0.011128892190754414, + 0.012382668443024158, + 0.012545688077807426, + 0.0083040501922369, + 0.007452450226992369, + 0.01117695402354002, + 0.011171450838446617, + 0.010275185108184814, + 0.010648387484252453, + 0.036457281559705734, + 0.02243928238749504, + 0.003964090719819069, + 0.0030470441561192274, + 0.006032556761056185, + 0.0063339476473629475, + 0.004413450602442026, + 0.0059068030677735806, + 0.012630839832127094, + 0.009186025708913803, + 0.0033732480369508266, + 0.0024689477868378162, + 0.004988220054656267, + 0.004128920380026102, + 0.0037821936421096325, + 0.0036297044716775417, + 0.0037411446683108807, + 0.004740242846310139, + 0.007083751726895571, + 0.007775061298161745, + 0.009028695523738861, + 0.009693230502307415, + 0.016157539561390877, + 0.01051357202231884, + 0.017329467460513115, + 0.015175753273069859, + 0.006421797908842564, + 0.007770055439323187, + 0.007915019989013672, + 0.00689851026982069, + 0.0185873880982399, + 0.009755616076290607, + 0.022120939567685127, + 0.01617203652858734, + 0.010337634943425655, + 0.008905657567083836, + 0.008860133588314056, + 0.008388800546526909, + 0.022520532831549644, + 0.015200121328234673, + 0.029256973415613174, + 0.021568773314356804, + 0.012874554842710495, + 0.01497503649443388, + 0.019564855843782425, + 0.01578284054994583, + 0.049504462629556656, + 0.02097100391983986, + 0.05342116206884384, + 0.04425838217139244, + 0.007417984772473574, + 0.010537541471421719, + 0.014329060912132263, + 0.011069530621170998, + 0.018437087535858154, + 0.011092049069702625, + 0.026292236521840096, + 0.021507419645786285, + 0.003936068620532751, + 0.004211151506751776, + 0.00706553366035223, + 0.008367273025214672, + 0.007198323495686054, + 0.005476885009557009, + 0.010281396098434925, + 0.009278198704123497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.01607622765004635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539346933364868 + ], + "opposite_logprob_tokens": [ + -0.42893466353416443 + ], + "logprob_reward": -1.0539346933364868, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539346933364868, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1611183136701584, + 0.08218565583229065, + 0.055325575172901154, + 0.03886529803276062, + 0.0366569422185421, + 0.05452239513397217, + 0.05898456647992134, + 0.17616556584835052, + 0.13949348032474518, + 0.19668222963809967 + ], + "credit_map": [ + 0.01093932893127203, + 0.01796899177134037, + 0.022385593503713608, + 0.015560432337224483, + 0.019312992691993713, + 0.020992962643504143, + 0.027970006689429283, + 0.025988005101680756, + 0.006241553463041782, + 0.00649945018813014, + 0.010205931030213833, + 0.00790404248982668, + 0.008543746545910835, + 0.007826133631169796, + 0.020542694255709648, + 0.014422104693949223, + 0.00494777038693428, + 0.004069313872605562, + 0.007646152749657631, + 0.007420955691486597, + 0.003819693112745881, + 0.006230917759239674, + 0.012363318353891373, + 0.0088274497538805, + 0.003601479809731245, + 0.0033209219109266996, + 0.0065494501031935215, + 0.003711898811161518, + 0.005192580167204142, + 0.0061942702159285545, + 0.0043550701811909676, + 0.005939628928899765, + 0.003992394544184208, + 0.0018771822797134519, + 0.006252211052924395, + 0.0035065803676843643, + 0.003134408500045538, + 0.004136086907237768, + 0.008003639988601208, + 0.005754436831921339, + 0.003908271435648203, + 0.0039665428921580315, + 0.0048264977522194386, + 0.004225875250995159, + 0.008682098239660263, + 0.006061593536287546, + 0.013705629855394363, + 0.009145884774625301, + 0.003879536408931017, + 0.0035257020499557257, + 0.0037080200854688883, + 0.004174060188233852, + 0.005998197942972183, + 0.004712565802037716, + 0.020968345925211906, + 0.012018142268061638, + 0.010461513884365559, + 0.01242613885551691, + 0.015333862975239754, + 0.013217994943261147, + 0.030638957396149635, + 0.01781773939728737, + 0.04347149655222893, + 0.03279786556959152, + 0.010142983868718147, + 0.01818416826426983, + 0.014399440959095955, + 0.012273524887859821, + 0.01997554302215576, + 0.01877095177769661, + 0.022072436287999153, + 0.02367442473769188, + 0.009261962957680225, + 0.011813011020421982, + 0.013579859398305416, + 0.012908935546875, + 0.021714597940444946, + 0.014924668706953526, + 0.05504002049565315, + 0.057439178228378296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -1.0539346933364868 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, static shot focuses with macroscopic precision on a translucent icicle suspended from the weathered, dark wooden ledge of an exterior window frame. The low, golden light of the winter afternoon sun casts a shadow on the icicle, revealing its crystalline structure and the fine frost on its surface. At its sharp tip, a perfectly spherical droplet of meltwater deliberately expands, its surface tension creating a clear, inverted, blurred snowscape in the background. Under its own weight, the droplet stretches downward, its shape elongating from a sphere to a teardrop. The thin thread connecting it to the icicle shimmers for a moment before inevitably breaking, and the droplet quickly and silently falls from the bottom of the frame, temporarily dampening the tip of the icicle. The scene is imbued with a tranquil, quiet atmosphere, capturing the subtle transitional moment with realistic detail.", + "seed": 24336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_14/index_9643_seed_24336_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2456674669869244, + "vlm_logprob": -0.2456674669869244, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.2456674669869244, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0478939451277256 + ], + "opposite_logprob_tokens": [ + -4.04789400100708 + ], + "logprob_reward": -0.0478939451277256, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0478939451277256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08260584622621536, + 0.023705530911684036, + 0.06520318984985352, + 0.1161453053355217, + 0.05585289001464844, + 0.07958076149225235, + 0.14541786909103394, + 0.06291943788528442, + 0.3022255301475525, + 0.06634366512298584 + ], + "credit_map": [ + 0.006464204750955105, + 0.006272348575294018, + 0.012469826266169548, + 0.018420901149511337, + 0.008545695804059505, + 0.005766305141150951, + 0.015835609287023544, + 0.008830953389406204, + 0.0018916602712124586, + 0.002669496228918433, + 0.003643292235210538, + 0.004040490835905075, + 0.0026633755769580603, + 0.002013920107856393, + 0.0040343767032027245, + 0.0027489184867590666, + 0.005655109882354736, + 0.005825693253427744, + 0.010939001105725765, + 0.011850175447762012, + 0.0060600102879107, + 0.006195637863129377, + 0.011792577803134918, + 0.006884981878101826, + 0.01038189884275198, + 0.010880602523684502, + 0.016809865832328796, + 0.021102121099829674, + 0.013096491806209087, + 0.0116761215031147, + 0.020287737250328064, + 0.011910462751984596, + 0.009729146026074886, + 0.0038112658075988293, + 0.007120820228010416, + 0.007187230512499809, + 0.004557538311928511, + 0.0062465001828968525, + 0.009833409450948238, + 0.007366981357336044, + 0.011149613186717033, + 0.006541125476360321, + 0.012017701752483845, + 0.008773187175393105, + 0.0074413795955479145, + 0.009094355627894402, + 0.015421957708895206, + 0.009141444228589535, + 0.01552135031670332, + 0.013251129537820816, + 0.021879632025957108, + 0.02105194516479969, + 0.014992914162576199, + 0.01574961096048355, + 0.026759063825011253, + 0.016212226822972298, + 0.007536469027400017, + 0.006110311485826969, + 0.008138532750308514, + 0.006503603886812925, + 0.009534639306366444, + 0.00849931314587593, + 0.010538320057094097, + 0.006058253347873688, + 0.028166335076093674, + 0.04461037740111351, + 0.035976022481918335, + 0.043053243309259415, + 0.04421959072351456, + 0.027334708720445633, + 0.050493594259023666, + 0.028371643275022507, + 0.003834455506876111, + 0.00588741060346365, + 0.009371860884130001, + 0.010362586937844753, + 0.010202753357589245, + 0.0046636564657092094, + 0.015814032405614853, + 0.006206909194588661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.0478939451277256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00283577898517251 + ], + "opposite_logprob_tokens": [ + -5.877835750579834 + ], + "logprob_reward": -0.00283577898517251, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00283577898517251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13214148581027985, + 0.056852538138628006, + 0.057702720165252686, + 0.09000515192747116, + 0.11256235837936401, + 0.19274888932704926, + 0.08582980185747147, + 0.06337761133909225, + 0.05993271246552467, + 0.14884668588638306 + ], + "credit_map": [ + 0.011956869624555111, + 0.014289761893451214, + 0.018423154950141907, + 0.021883299574255943, + 0.01764836348593235, + 0.010536689311265945, + 0.025854147970676422, + 0.01154920645058155, + 0.004403518978506327, + 0.006359957158565521, + 0.00856857281178236, + 0.007420150097459555, + 0.0069117313250899315, + 0.004922154359519482, + 0.01242755725979805, + 0.00583889614790678, + 0.004665348678827286, + 0.00403845077380538, + 0.008504052646458149, + 0.01005393173545599, + 0.0055199116468429565, + 0.005681725684553385, + 0.01247002650052309, + 0.006769269704818726, + 0.0072971307672560215, + 0.007593969348818064, + 0.014545205980539322, + 0.015563547611236572, + 0.008579657413065434, + 0.008944354020059109, + 0.018070846796035767, + 0.009410439059138298, + 0.010933696292340755, + 0.011576801538467407, + 0.014655754901468754, + 0.01963820867240429, + 0.010822458192706108, + 0.012210707180202007, + 0.01931573823094368, + 0.01340899895876646, + 0.024046599864959717, + 0.016239915043115616, + 0.026821592822670937, + 0.02577078901231289, + 0.019695838913321495, + 0.01822509802877903, + 0.03624236211180687, + 0.025706708431243896, + 0.00928555428981781, + 0.006781649775803089, + 0.011759798973798752, + 0.013127673417329788, + 0.009070774540305138, + 0.008732503280043602, + 0.01508963480591774, + 0.011982218362390995, + 0.007219122722744942, + 0.007070701103657484, + 0.007664780132472515, + 0.0058261938393116, + 0.014289402402937412, + 0.006101314444094896, + 0.009493463672697544, + 0.005712630692869425, + 0.0058592986315488815, + 0.00614093616604805, + 0.008202495984733105, + 0.008380064740777016, + 0.008383551612496376, + 0.0038068240974098444, + 0.013231683522462845, + 0.005927857011556625, + 0.013928357511758804, + 0.017188776284456253, + 0.02147802710533142, + 0.02663666196167469, + 0.0180465467274189, + 0.009287909604609013, + 0.029358187690377235, + 0.01292220875620842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.00283577898517251 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355715468525887 + ], + "opposite_logprob_tokens": [ + -3.651355743408203 + ], + "logprob_reward": -0.026355715468525887, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355715468525887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10730019211769104, + 0.1504325419664383, + 0.0364157035946846, + 0.1303139179944992, + 0.08193785697221756, + 0.12598998844623566, + 0.10316799581050873, + 0.07093276083469391, + 0.11884187161922455, + 0.07466714829206467 + ], + "credit_map": [ + 0.01098905224353075, + 0.010006994940340519, + 0.015729669481515884, + 0.011684042401611805, + 0.018032608553767204, + 0.010539301671087742, + 0.019084393978118896, + 0.01123413909226656, + 0.010039783082902431, + 0.014555941335856915, + 0.0224397461861372, + 0.02236662060022354, + 0.018598543480038643, + 0.013899344950914383, + 0.03239959105849266, + 0.016132976859807968, + 0.0033607338555157185, + 0.002632091287523508, + 0.004911346361041069, + 0.005823993124067783, + 0.004002481233328581, + 0.003752939635887742, + 0.007593892980366945, + 0.004338224418461323, + 0.013343242928385735, + 0.011578008532524109, + 0.01792006753385067, + 0.02366272546350956, + 0.012276546098291874, + 0.015747662633657455, + 0.021011393517255783, + 0.014774279668927193, + 0.013881461694836617, + 0.00579630583524704, + 0.012365056201815605, + 0.012751694768667221, + 0.007099952083081007, + 0.010337335988879204, + 0.011662820354104042, + 0.008043232373893261, + 0.01307619921863079, + 0.013096310198307037, + 0.016578232869505882, + 0.013238172978162766, + 0.01641659438610077, + 0.014509288594126701, + 0.02532034181058407, + 0.013754844665527344, + 0.009464322589337826, + 0.00926161464303732, + 0.016147689893841743, + 0.015421725809574127, + 0.012206791900098324, + 0.010331560857594013, + 0.021236835047602654, + 0.009097455069422722, + 0.007978147827088833, + 0.005669708829373121, + 0.009654895402491093, + 0.010015309788286686, + 0.008272434584796429, + 0.008206966333091259, + 0.012085252441465855, + 0.009050044231116772, + 0.014134489931166172, + 0.012341929599642754, + 0.01316240057349205, + 0.011534362100064754, + 0.020642658695578575, + 0.01199368480592966, + 0.02244403213262558, + 0.012588316574692726, + 0.004074644297361374, + 0.005050886422395706, + 0.012420469895005226, + 0.013916720636188984, + 0.008046295493841171, + 0.004037535283714533, + 0.019582383334636688, + 0.00753821711987257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.026355715468525887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194829389452934 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014194829389452934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194829389452934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11534547805786133, + 0.06473111361265182, + 0.04605626314878464, + 0.1268111616373062, + 0.12176194041967392, + 0.1446484476327896, + 0.04809103533625603, + 0.06614848226308823, + 0.1513143926858902, + 0.11509175598621368 + ], + "credit_map": [ + 0.009481494314968586, + 0.012346596457064152, + 0.016657669097185135, + 0.019116109237074852, + 0.016911933198571205, + 0.010441703721880913, + 0.016286659985780716, + 0.01410331204533577, + 0.005635259207338095, + 0.005704863462597132, + 0.009338145144283772, + 0.008624370209872723, + 0.007369486149400473, + 0.006250890903174877, + 0.013271226547658443, + 0.0085368687286973, + 0.0030678946059197187, + 0.0032989683095365763, + 0.007489094045013189, + 0.009165795519948006, + 0.004212336614727974, + 0.0049696629866957664, + 0.00974402017891407, + 0.004108487628400326, + 0.013185175135731697, + 0.010467403568327427, + 0.017527908086776733, + 0.02272452600300312, + 0.01465355884283781, + 0.01460215263068676, + 0.020929666236042976, + 0.01272076740860939, + 0.01810821332037449, + 0.009634089656174183, + 0.01710154488682747, + 0.02052021399140358, + 0.010922055691480637, + 0.014227419160306454, + 0.017254043370485306, + 0.013994358479976654, + 0.015622741542756557, + 0.01159315463155508, + 0.021094918251037598, + 0.019765425473451614, + 0.01701383665204048, + 0.01320072915405035, + 0.029995424672961235, + 0.01636221632361412, + 0.004741298034787178, + 0.004542798735201359, + 0.006653079763054848, + 0.006498726084828377, + 0.006572106387466192, + 0.004375942517071962, + 0.010192414745688438, + 0.00451466953381896, + 0.008005782961845398, + 0.006290844175964594, + 0.008956044912338257, + 0.008022570982575417, + 0.0097673200070858, + 0.00781932007521391, + 0.010187901556491852, + 0.007098699454218149, + 0.015356488525867462, + 0.016794197261333466, + 0.017954517155885696, + 0.016187358647584915, + 0.027494952082633972, + 0.01619790866971016, + 0.028798962011933327, + 0.012530013918876648, + 0.008088937029242516, + 0.010251489467918873, + 0.014644517563283443, + 0.020727647468447685, + 0.01893419772386551, + 0.0066666509956121445, + 0.028019912540912628, + 0.0077584004029631615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a translucent icicle suspended from a weathered, dark wooden ledge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly spherical droplet of meltwater expand at the sharp tip of the icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet stretch from a sphere into a teardrop shape before the connecting thread breaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the droplet fall from the bottom of the frame and temporarily dampen the tip of the icicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370570659637451 + ], + "opposite_logprob_tokens": [ + -0.3870570659637451 + ], + "logprob_reward": -1.1370570659637451, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370570659637451, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04147299379110336, + 0.018252234905958176, + 0.028790531679987907, + 0.045161422342061996, + 0.06431576609611511, + 0.03809116408228874, + 0.09776564687490463, + 0.10130517929792404, + 0.3876792788505554, + 0.17716577649116516 + ], + "credit_map": [ + 0.0046661412343382835, + 0.0040822080336511135, + 0.00729071069508791, + 0.005825840402394533, + 0.006427014246582985, + 0.003844615537673235, + 0.0056099616922438145, + 0.0037265003193169832, + 0.0016620215028524399, + 0.0014201385201886296, + 0.002726296428591013, + 0.0026096391957253218, + 0.0019331341609358788, + 0.0016958591295406222, + 0.0037394387181848288, + 0.002465709112584591, + 0.002226254204288125, + 0.0029328942764550447, + 0.004535046871751547, + 0.005308791529387236, + 0.0031053393613547087, + 0.0021949606016278267, + 0.005748990923166275, + 0.002738253679126501, + 0.002889004535973072, + 0.0040430063381791115, + 0.007633183151483536, + 0.008031059056520462, + 0.004501333925873041, + 0.0038004652597010136, + 0.008886869065463543, + 0.0053765010088682175, + 0.008436651900410652, + 0.006271476857364178, + 0.0077974130399525166, + 0.010481002740561962, + 0.00594508508220315, + 0.006563163828104734, + 0.011426174081861973, + 0.007394794840365648, + 0.004494867287576199, + 0.003176906146109104, + 0.006191574037075043, + 0.005000157747417688, + 0.00289913360029459, + 0.002852641511708498, + 0.008872690610587597, + 0.004603193607181311, + 0.00955240149050951, + 0.008378843776881695, + 0.014829153195023537, + 0.014060788787901402, + 0.011451838538050652, + 0.008997506462037563, + 0.022717580199241638, + 0.00777753721922636, + 0.012945457361638546, + 0.00922662578523159, + 0.013713525608181953, + 0.01017195824533701, + 0.014640075154602528, + 0.013637064956128597, + 0.015976164489984512, + 0.010994302108883858, + 0.037041421979665756, + 0.05779784172773361, + 0.04546545073390007, + 0.05498884990811348, + 0.05711974576115608, + 0.03586669638752937, + 0.06326208263635635, + 0.036137185990810394, + 0.008847544901072979, + 0.015248249284923077, + 0.02960950881242752, + 0.028192952275276184, + 0.020792797207832336, + 0.009673546999692917, + 0.049768269062042236, + 0.015032907016575336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -1.1370570659637451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_tensor([5961, 5961])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_tensor([5961, 5961])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90a88d8a865bafd5752926931ac52231d889238b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_tensor([5961, 5961])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene begins with a slow, deliberate left-to-right pan, initiated by a brilliant, anamorphic lens flare that washes the frame in streaks of warm gold and ethereal blue light. As the flash gracefully sweeps across the lens, it reveals a low-angle shot of a truly enormous python, calmly coiled around a sun-drenched, mossy boulder in a humid jungle clearing. The snake's massive body is wound into a tight, intricate spiral, its scales a captivating iridescent mosaic of olive green and rich earthy brown, forming an intricate diamond pattern that shimmers in the intense sunlight. Each individual scale is so distinct that its texture and the way it catches the light are visible. The creature is in a state of deep rest; its head is tucked into its coil, the only perceptible movement being the slow, rhythmic expansion and contraction of its powerful abdomen as it breathes. The environment is lush and vibrant, with golden afternoon sunlight penetrating the dense jungle canopy above, dotting the forest floor and the snake's shimmering back with spots. The background is a soft-focus tapestry of deep green ferns, hanging vines, and the dark, fertile soil of the jungle. The overall atmosphere is one of profound tranquility and majestic power, capturing a quiet, awe-inspiring moment of nature in the warm, hazy glow of the sun.", + "seed": 31248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31248_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06613530479371547, + "vlm_logprob": -0.06613530479371547, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06613530479371547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1055951714515686 + ], + "opposite_logprob_tokens": [ + -2.355595111846924 + ], + "logprob_reward": -0.1055951714515686, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1055951714515686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06255365908145905, + 0.01931343786418438, + 0.0635925903916359, + 0.0874050036072731, + 0.024068759754300117, + 0.042231373488903046, + 0.03636855632066727, + 0.5282170176506042, + 0.0635591447353363, + 0.07269056141376495 + ], + "credit_map": [ + 0.005522331688553095, + 0.00792421493679285, + 0.008631084114313126, + 0.010556459426879883, + 0.008128339424729347, + 0.0064749387092888355, + 0.006746670231223106, + 0.00856961589306593, + 0.0015896728727966547, + 0.0027062383014708757, + 0.0018178545869886875, + 0.00270694587379694, + 0.0036598604638129473, + 0.0021456987597048283, + 0.0023683293256908655, + 0.002318835351616144, + 0.0036675152368843555, + 0.007815971970558167, + 0.005767069291323423, + 0.00936626922339201, + 0.006757184397429228, + 0.009200610220432281, + 0.009542453102767467, + 0.011475512757897377, + 0.006170718465000391, + 0.005940703209489584, + 0.009477374143898487, + 0.010118882171809673, + 0.019556326791644096, + 0.014433245174586773, + 0.013814843259751797, + 0.007892905734479427, + 0.0018087272765114903, + 0.0019582631066441536, + 0.003002085257321596, + 0.0025575372856110334, + 0.003593952627852559, + 0.0036433737259358168, + 0.004113932605832815, + 0.0033908856566995382, + 0.004209305625408888, + 0.002759983530268073, + 0.004260202869772911, + 0.004816124215722084, + 0.007102761883288622, + 0.006862461566925049, + 0.006262396927922964, + 0.005958132445812225, + 0.004677517805248499, + 0.0032490112353116274, + 0.005347680300474167, + 0.00406286958605051, + 0.0056790877133607864, + 0.004094830714166164, + 0.0047949934378266335, + 0.004462564364075661, + 0.060618314892053604, + 0.012277308851480484, + 0.11936568468809128, + 0.033567819744348526, + 0.0365150049328804, + 0.20132479071617126, + 0.02504834160208702, + 0.0394996777176857, + 0.009445816278457642, + 0.009110906161367893, + 0.00800467748194933, + 0.005582224577665329, + 0.004817091394215822, + 0.012724973261356354, + 0.00695800269022584, + 0.006915445439517498, + 0.010458076372742653, + 0.010122153908014297, + 0.009089941158890724, + 0.009690066799521446, + 0.006713836453855038, + 0.008949809707701206, + 0.008451736532151699, + 0.009214934892952442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.1055951714515686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428530514240265 + ], + "opposite_logprob_tokens": [ + -2.017853021621704 + ], + "logprob_reward": -0.1428530514240265, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428530514240265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26637911796569824, + 0.07776009291410446, + 0.11814408004283905, + 0.10061227530241013, + 0.05614149197936058, + 0.06297454982995987, + 0.015423394739627838, + 0.24337182939052582, + 0.026384487748146057, + 0.03280863165855408 + ], + "credit_map": [ + 0.023690741509199142, + 0.05160237476229668, + 0.019056005403399467, + 0.039058852940797806, + 0.03839917480945587, + 0.029047507792711258, + 0.03467989340424538, + 0.03084457479417324, + 0.005452489480376244, + 0.008195079863071442, + 0.009193293750286102, + 0.01282808743417263, + 0.011828165501356125, + 0.00928665790706873, + 0.011745893396437168, + 0.009230424650013447, + 0.0065400563180446625, + 0.01644054986536503, + 0.010229632258415222, + 0.013745819218456745, + 0.015734538435935974, + 0.01859079673886299, + 0.01453177910298109, + 0.022330911830067635, + 0.007769125048071146, + 0.007251776289194822, + 0.012429989874362946, + 0.008782901801168919, + 0.024337414652109146, + 0.016808895394206047, + 0.013346953317523003, + 0.009885218925774097, + 0.0037671695463359356, + 0.003969515673816204, + 0.004178633447736502, + 0.006844346411526203, + 0.009858055971562862, + 0.007457748055458069, + 0.011976801790297031, + 0.008089223876595497, + 0.004807996563613415, + 0.0026683027390390635, + 0.003279331838712096, + 0.005197626538574696, + 0.007944857701659203, + 0.01759592443704605, + 0.006099908612668514, + 0.015380602329969406, + 0.002697769785299897, + 0.001093026134185493, + 0.0015846575843170285, + 0.0011813952587544918, + 0.0039014280773699284, + 0.002027086215093732, + 0.001403913600370288, + 0.0015341187827289104, + 0.026371847838163376, + 0.006230010651051998, + 0.060350384563207626, + 0.017066840082406998, + 0.01569260284304619, + 0.08854776620864868, + 0.011907147243618965, + 0.017205238342285156, + 0.003866618499159813, + 0.0046457466669380665, + 0.0036003305576741695, + 0.002170681022107601, + 0.002042782958596945, + 0.004046970512717962, + 0.002552743535488844, + 0.0034586123656481504, + 0.004642886575311422, + 0.004850103985518217, + 0.004849348217248917, + 0.003285734448581934, + 0.00311859929934144, + 0.003597484203055501, + 0.0039473543874919415, + 0.00451712217181921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.1428530514240265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100250080227852 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100250080227852, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100250080227852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06674493104219437, + 0.017351748421788216, + 0.07478298991918564, + 0.07971041649580002, + 0.06859026104211807, + 0.05276467278599739, + 0.04827985540032387, + 0.5503868460655212, + 0.026010092347860336, + 0.015378170646727085 + ], + "credit_map": [ + 0.00629129633307457, + 0.010421836748719215, + 0.005969171412289143, + 0.00873326975852251, + 0.009074965491890907, + 0.00685206800699234, + 0.010994651354849339, + 0.008407683111727238, + 0.0012581328628584743, + 0.0028134521562606096, + 0.002059789141640067, + 0.0026753146667033434, + 0.0035568219609558582, + 0.0016819869633764029, + 0.0019194483757019043, + 0.0013868032256141305, + 0.004413980524986982, + 0.009249664843082428, + 0.007772366516292095, + 0.017118681222200394, + 0.008299396373331547, + 0.007890010252594948, + 0.012541781179606915, + 0.007497111801058054, + 0.004263246897608042, + 0.006968117319047451, + 0.00810722354799509, + 0.006899615284055471, + 0.01972564309835434, + 0.017529891803860664, + 0.009990883059799671, + 0.0062258001416921616, + 0.0036838743835687637, + 0.0038293120451271534, + 0.005950080696493387, + 0.008272725157439709, + 0.008284684270620346, + 0.011372596956789494, + 0.016192080453038216, + 0.011004908010363579, + 0.005292069632560015, + 0.002957537304610014, + 0.0048059881664812565, + 0.005290066823363304, + 0.00746539793908596, + 0.011736202985048294, + 0.005862799007445574, + 0.009354619309306145, + 0.007453208789229393, + 0.003993048332631588, + 0.00597827835008502, + 0.004362625535577536, + 0.009381763637065887, + 0.005858283024281263, + 0.0057241362519562244, + 0.005528515204787254, + 0.06155836209654808, + 0.012785824947059155, + 0.1305173933506012, + 0.03707772120833397, + 0.03687075525522232, + 0.2059977799654007, + 0.026049625128507614, + 0.03952942416071892, + 0.004157202783972025, + 0.003216053592041135, + 0.003462425200268626, + 0.003663198323920369, + 0.002384850289672613, + 0.003999852109700441, + 0.0023008205462247133, + 0.002825693693011999, + 0.0017586909234523773, + 0.001464762375690043, + 0.003132354002445936, + 0.0022867871448397636, + 0.0010757140116766095, + 0.0016174732008948922, + 0.0020132146310061216, + 0.0020291758701205254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.004100250080227852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075171530246735 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075171530246735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075171530246735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09762836992740631, + 0.02154141291975975, + 0.018369803205132484, + 0.03671742603182793, + 0.05346810817718506, + 0.06898654252290726, + 0.032928239554166794, + 0.6270001530647278, + 0.031143933534622192, + 0.012215961702167988 + ], + "credit_map": [ + 0.009596733376383781, + 0.016209082677960396, + 0.01046889740973711, + 0.018503867089748383, + 0.01340106874704361, + 0.009825771674513817, + 0.010108268819749355, + 0.009514674544334412, + 0.00154106505215168, + 0.0027714879252016544, + 0.0029684973414987326, + 0.003925454802811146, + 0.0029389122501015663, + 0.002087186323478818, + 0.002853330923244357, + 0.0024554773699492216, + 0.001127477502450347, + 0.002126439241692424, + 0.0024324164260178804, + 0.0026322833728045225, + 0.0021242559887468815, + 0.0024555078707635403, + 0.0032891370356082916, + 0.0021822871640324593, + 0.0022611457388848066, + 0.0021125846542418003, + 0.0057417587377130985, + 0.0038465987890958786, + 0.006895966362208128, + 0.005528133362531662, + 0.006571390200406313, + 0.0037598474882543087, + 0.004575059283524752, + 0.003498150035738945, + 0.007149058394134045, + 0.005025195889174938, + 0.010069397278130054, + 0.007166952360421419, + 0.008443656377494335, + 0.0075406357645988464, + 0.0063833449967205524, + 0.0032539633102715015, + 0.004686321597546339, + 0.004421905614435673, + 0.008520739153027534, + 0.020078333094716072, + 0.006265260744839907, + 0.015376678667962551, + 0.004661307670176029, + 0.0028769029304385185, + 0.0044393292628228664, + 0.0030043581500649452, + 0.005215346347540617, + 0.004658679477870464, + 0.003889172337949276, + 0.004183141048997641, + 0.07037845253944397, + 0.014977340586483479, + 0.1476154774427414, + 0.04153317213058472, + 0.04198189824819565, + 0.2354465126991272, + 0.029606403782963753, + 0.04546092823147774, + 0.0037629122380167246, + 0.006366791669279337, + 0.0038491804152727127, + 0.002181987278163433, + 0.002886170521378517, + 0.005901156924664974, + 0.0028946769889444113, + 0.0033010588958859444, + 0.0015967156505212188, + 0.0015668361447751522, + 0.0017854569014161825, + 0.0013421372277662158, + 0.001034726039506495, + 0.0017834982136264443, + 0.0014910000609233975, + 0.001615591230802238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.016075171530246735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062052879482507706 + ], + "opposite_logprob_tokens": [ + -2.8120529651641846 + ], + "logprob_reward": -0.062052879482507706, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062052879482507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2758103311061859, + 0.036094360053539276, + 0.06948941946029663, + 0.10811033099889755, + 0.10360503196716309, + 0.09215246140956879, + 0.13392391800880432, + 0.05177360773086548, + 0.10487134009599686, + 0.024169184267520905 + ], + "credit_map": [ + 0.03135576844215393, + 0.02995632216334343, + 0.030626805499196053, + 0.07370490580797195, + 0.02839568629860878, + 0.01630629412829876, + 0.041652120649814606, + 0.02381242997944355, + 0.002576138125732541, + 0.004212965723127127, + 0.004040585830807686, + 0.005162016488611698, + 0.005322966258972883, + 0.004852091893553734, + 0.0052238088101148605, + 0.004703785292804241, + 0.0054838452488183975, + 0.010384800843894482, + 0.007402611896395683, + 0.008526066318154335, + 0.008123520761728287, + 0.010042057372629642, + 0.00788049679249525, + 0.011646016500890255, + 0.006456267554312944, + 0.010359796695411205, + 0.01036711223423481, + 0.011002011597156525, + 0.02815868705511093, + 0.018959131091833115, + 0.015137048438191414, + 0.007670277263969183, + 0.007627449929714203, + 0.006598885636776686, + 0.008139386773109436, + 0.014666871167719364, + 0.015023878775537014, + 0.01451116893440485, + 0.02196674607694149, + 0.01507064700126648, + 0.009690141305327415, + 0.00661330483853817, + 0.010773982852697372, + 0.010744040831923485, + 0.016125768423080444, + 0.011538638733327389, + 0.014312073588371277, + 0.012354516424238682, + 0.018753360956907272, + 0.010731680318713188, + 0.019429108127951622, + 0.012003005482256413, + 0.022778965532779694, + 0.018345925956964493, + 0.01611321046948433, + 0.015768663957715034, + 0.007499143946915865, + 0.006199672818183899, + 0.006323985755443573, + 0.0064808037132024765, + 0.004406836349517107, + 0.008033367805182934, + 0.007234838325530291, + 0.005594955757260323, + 0.013139340095221996, + 0.02215149812400341, + 0.013239522464573383, + 0.009650408290326595, + 0.00872656237334013, + 0.017261534929275513, + 0.007740627974271774, + 0.012961842119693756, + 0.002470505191013217, + 0.0036429977044463158, + 0.004359671846032143, + 0.0022165910340845585, + 0.0022789696231484413, + 0.0029577987734228373, + 0.002475124318152666, + 0.003767527872696519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.062052879482507706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, deliberate left-to-right pan, initiated by a brilliant, anamorphic lens flare that washes the frame in streaks of warm gold and ethereal blue light. As the flash gracefully sweeps across the lens, it reveals a low-angle shot of a truly enormous python, calmly coiled around a sun-drenched, mossy boulder in a humid jungle clearing. The snake's massive body is wound into a tight, intricate spiral, its scales a captivating iridescent mosaic of olive green and rich earthy brown, forming an intricate diamond pattern that shimmers in the intense sunlight. Each individual scale is so distinct that its texture and the way it catches the light are visible. The creature is in a state of deep rest; its head is tucked into its coil, the only perceptible movement being the slow, rhythmic expansion and contraction of its powerful abdomen as it breathes. The environment is lush and vibrant, with golden afternoon sunlight penetrating the dense jungle canopy above, dotting the forest floor and the snake's shimmering back with spots. The background is a soft-focus tapestry of deep green ferns, hanging vines, and the dark, fertile soil of the jungle. The overall atmosphere is one of profound tranquility and majestic power, capturing a quiet, awe-inspiring moment of nature in the warm, hazy glow of the sun.", + "seed": 31249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31249_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05064785862341523, + "vlm_logprob": -0.05064785862341523, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05064785862341523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13447237014770508 + ], + "opposite_logprob_tokens": [ + -2.134472370147705 + ], + "logprob_reward": -0.13447237014770508, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13447237014770508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35677245259284973, + 0.038256749510765076, + 0.23353923857212067, + 0.042623162269592285, + 0.06512695550918579, + 0.023264393210411072, + 0.13544638454914093, + 0.024129049852490425, + 0.03941907361149788, + 0.0414225235581398 + ], + "credit_map": [ + 0.04387185722589493, + 0.04004320129752159, + 0.04940188303589821, + 0.06613712757825851, + 0.06521297991275787, + 0.041747450828552246, + 0.029088404029607773, + 0.02126959152519703, + 0.002773879561573267, + 0.004229420330375433, + 0.004957565572112799, + 0.006353938486427069, + 0.00586296571418643, + 0.0062225437723100185, + 0.004941247869282961, + 0.002915189368650317, + 0.02728406898677349, + 0.03609248623251915, + 0.018495706841349602, + 0.03226573020219803, + 0.027859169989824295, + 0.0445542074739933, + 0.029187798500061035, + 0.017800109460949898, + 0.004021544009447098, + 0.0047821602784097195, + 0.0033788857981562614, + 0.0074782236479222775, + 0.003927646670490503, + 0.008544606156647205, + 0.00426224572584033, + 0.006227853242307901, + 0.0069322544150054455, + 0.006495967041701078, + 0.004631262272596359, + 0.009095444343984127, + 0.0059119644574820995, + 0.009642067365348339, + 0.009902812540531158, + 0.012515191920101643, + 0.002069031586870551, + 0.002432253211736679, + 0.0016700716223567724, + 0.002361116698011756, + 0.002848994452506304, + 0.0038388839457184076, + 0.004066411405801773, + 0.003977633081376553, + 0.008465416729450226, + 0.00605746591463685, + 0.01240228395909071, + 0.018296461552381516, + 0.009219995699822903, + 0.010444075800478458, + 0.03956031799316406, + 0.031000381335616112, + 0.0028909179382026196, + 0.00207977881655097, + 0.002512144623324275, + 0.002882149536162615, + 0.0026162273716181517, + 0.0032178794499486685, + 0.004552742000669241, + 0.0033772115129977465, + 0.005328946281224489, + 0.0036890816409140825, + 0.005364762619137764, + 0.003484459826722741, + 0.003412243677303195, + 0.005915797781199217, + 0.006381392944604158, + 0.005842389538884163, + 0.0035907875280827284, + 0.003301462857052684, + 0.007293473929166794, + 0.005014199297875166, + 0.004541815724223852, + 0.006656923331320286, + 0.005724428221583366, + 0.005299438256770372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.13447237014770508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984067052602768 + ], + "opposite_logprob_tokens": [ + -3.5298407077789307 + ], + "logprob_reward": -0.02984067052602768, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984067052602768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2891892194747925, + 0.1826649308204651, + 0.06133197993040085, + 0.0797567367553711, + 0.08233879506587982, + 0.027698522433638573, + 0.13930779695510864, + 0.035670261830091476, + 0.04906372353434563, + 0.052978068590164185 + ], + "credit_map": [ + 0.0339662991464138, + 0.03196650370955467, + 0.036387789994478226, + 0.05766282230615616, + 0.03757447004318237, + 0.03670733422040939, + 0.03596779331564903, + 0.018956167623400688, + 0.016843434423208237, + 0.022286193445324898, + 0.02283606305718422, + 0.028719257563352585, + 0.02467530593276024, + 0.02378344163298607, + 0.025670256465673447, + 0.017850952222943306, + 0.00654171546921134, + 0.009625532664358616, + 0.0044775367714464664, + 0.008852812461555004, + 0.007525885012000799, + 0.011215273290872574, + 0.008142538368701935, + 0.004950681235641241, + 0.00721374386921525, + 0.011726179160177708, + 0.010369467549026012, + 0.011053149588406086, + 0.008422010578215122, + 0.013510075397789478, + 0.007742955815047026, + 0.00971914455294609, + 0.011007547378540039, + 0.009113573469221592, + 0.005798083730041981, + 0.008678856305778027, + 0.0089418925344944, + 0.011742263101041317, + 0.014619178138673306, + 0.01243739016354084, + 0.002834287006407976, + 0.0035769501700997353, + 0.0021881875582039356, + 0.003786042332649231, + 0.003849887754768133, + 0.0032942788675427437, + 0.003707422874867916, + 0.004461462143808603, + 0.0043501355685293674, + 0.005620851181447506, + 0.0065314192324876785, + 0.011155031621456146, + 0.00827151071280241, + 0.009230000898241997, + 0.04956389218568802, + 0.04458494484424591, + 0.0044194720685482025, + 0.004238399211317301, + 0.004003054462373257, + 0.004048243630677462, + 0.004190504085272551, + 0.004871791694313288, + 0.005120605695992708, + 0.004778187721967697, + 0.004925833083689213, + 0.00562237948179245, + 0.005649990867823362, + 0.004607168026268482, + 0.005226468667387962, + 0.009417536668479443, + 0.006738788913935423, + 0.00687554944306612, + 0.006278904154896736, + 0.006349119823426008, + 0.007393228821456432, + 0.0067520118318498135, + 0.00484834797680378, + 0.008652755059301853, + 0.005226398352533579, + 0.007477296981960535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02984067052602768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630605716258287 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630605716258287, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630605716258287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07788863778114319, + 0.140542671084404, + 0.02283535711467266, + 0.03869360685348511, + 0.07874242216348648, + 0.0832507461309433, + 0.21447694301605225, + 0.09911749511957169, + 0.09650308638811111, + 0.14794902503490448 + ], + "credit_map": [ + 0.009496553801000118, + 0.007733630016446114, + 0.012539941817522049, + 0.011595366522669792, + 0.011019271798431873, + 0.010659905150532722, + 0.009765992872416973, + 0.0050779725424945354, + 0.012062910944223404, + 0.019122865051031113, + 0.017942549660801888, + 0.02429216541349888, + 0.021265141665935516, + 0.024668367579579353, + 0.0113618029281497, + 0.009826863184571266, + 0.0021605337969958782, + 0.0029079632367938757, + 0.001924764597788453, + 0.0038055810146033764, + 0.0022812914103269577, + 0.0037733917124569416, + 0.0034802898298949003, + 0.002501540584489703, + 0.00264224992133677, + 0.0033010023180395365, + 0.0023224165197461843, + 0.007504611741751432, + 0.004261733964085579, + 0.011213158257305622, + 0.003267197636887431, + 0.0041812327690422535, + 0.008561810478568077, + 0.007116687949746847, + 0.005337551701813936, + 0.009928095154464245, + 0.008827798999845982, + 0.009647971019148827, + 0.01678614132106304, + 0.012536358088254929, + 0.006237243767827749, + 0.006946085952222347, + 0.006232779007405043, + 0.010521912947297096, + 0.010574384592473507, + 0.015254229307174683, + 0.013039952144026756, + 0.014444153755903244, + 0.014473884366452694, + 0.014599804766476154, + 0.012121615931391716, + 0.013421899639070034, + 0.012107543647289276, + 0.014271406456828117, + 0.06929727643728256, + 0.06418349593877792, + 0.011430361308157444, + 0.005741807632148266, + 0.006963103543967009, + 0.008353159762918949, + 0.015056988224387169, + 0.0168574471026659, + 0.020302677527070045, + 0.014411945827305317, + 0.007648078259080648, + 0.006569392513483763, + 0.009760784916579723, + 0.010161908343434334, + 0.011117078363895416, + 0.01984524540603161, + 0.016185222193598747, + 0.015215369872748852, + 0.011110532097518444, + 0.011935101822018623, + 0.026694167405366898, + 0.017952004447579384, + 0.01621110551059246, + 0.01949131302535534, + 0.025445690378546715, + 0.019109109416604042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.007630605716258287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09509836882352829, + 0.08502237498760223, + 0.024416424334049225, + 0.01874384842813015, + 0.06481262296438217, + 0.07786334306001663, + 0.416579008102417, + 0.05146792158484459, + 0.03918498381972313, + 0.12681108713150024 + ], + "credit_map": [ + 0.010355903767049313, + 0.0090670520439744, + 0.011288092471659184, + 0.014526251703500748, + 0.017143163830041885, + 0.01105706486850977, + 0.0106661943718791, + 0.010994648560881615, + 0.008178754709661007, + 0.010367064736783504, + 0.010807022452354431, + 0.015732888132333755, + 0.011663714423775673, + 0.012863398529589176, + 0.00791224930435419, + 0.007497283164411783, + 0.0023139500990509987, + 0.003301838878542185, + 0.002399201039224863, + 0.0037192481104284525, + 0.0028350036591291428, + 0.0032519253436475992, + 0.003996607381850481, + 0.0025986479595303535, + 0.0013160574017092586, + 0.0016135437181219459, + 0.0017229419900104403, + 0.004035307094454765, + 0.0016377493739128113, + 0.002881573745980859, + 0.0024768260773271322, + 0.003059850540012121, + 0.006469337269663811, + 0.00747399078682065, + 0.005379366688430309, + 0.009363642893731594, + 0.008126266300678253, + 0.007326756604015827, + 0.012373159639537334, + 0.00830010138452053, + 0.00465820636600256, + 0.005213914439082146, + 0.005809598136693239, + 0.007778830826282501, + 0.008962802588939667, + 0.015748966485261917, + 0.01565222442150116, + 0.01403879839926958, + 0.01372640673071146, + 0.015882233157753944, + 0.025094734504818916, + 0.038452669978141785, + 0.02777927555143833, + 0.026160188019275665, + 0.14384472370147705, + 0.12563881278038025, + 0.005649734754115343, + 0.004919029306620359, + 0.004961132537573576, + 0.004177459049969912, + 0.005842329002916813, + 0.01059649046510458, + 0.007896841503679752, + 0.007424908224493265, + 0.004516514018177986, + 0.0036646253429353237, + 0.004389586858451366, + 0.004456689581274986, + 0.0041880239732563496, + 0.005855009891092777, + 0.005218101665377617, + 0.006896431557834148, + 0.008058718405663967, + 0.009337580762803555, + 0.01432858221232891, + 0.01605464145541191, + 0.015436619520187378, + 0.018245210871100426, + 0.0280733834952116, + 0.01727634109556675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494515597820282 + ], + "opposite_logprob_tokens": [ + -2.929945230484009 + ], + "logprob_reward": -0.05494515597820282, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494515597820282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13924777507781982, + 0.06092880666255951, + 0.030788391828536987, + 0.02744394913315773, + 0.06598381698131561, + 0.1021483764052391, + 0.10408341884613037, + 0.22764705121517181, + 0.08600714057683945, + 0.15572135150432587 + ], + "credit_map": [ + 0.01326733361929655, + 0.01938364841043949, + 0.020125865936279297, + 0.02323291078209877, + 0.024568866938352585, + 0.014755467884242535, + 0.010679273866117, + 0.013234397396445274, + 0.005817411467432976, + 0.008960229344666004, + 0.006323288660496473, + 0.00996220950037241, + 0.00967872329056263, + 0.007855131290853024, + 0.007566151209175587, + 0.004765654448419809, + 0.003056867979466915, + 0.003281446173787117, + 0.003533853217959404, + 0.004949646070599556, + 0.004401519428938627, + 0.003761605592444539, + 0.003946838900446892, + 0.0038566123694181442, + 0.0022204520646482706, + 0.0031394294928759336, + 0.0028403906617313623, + 0.005292678717523813, + 0.002861317014321685, + 0.005386881995946169, + 0.003180238651111722, + 0.002522560302168131, + 0.005457743536680937, + 0.00842369720339775, + 0.005723872687667608, + 0.00832305196672678, + 0.006633232347667217, + 0.009243634529411793, + 0.009831900708377361, + 0.012346680276095867, + 0.005719184875488281, + 0.00899452343583107, + 0.011808055453002453, + 0.00962577760219574, + 0.013672156259417534, + 0.016959642991423607, + 0.016831304877996445, + 0.018537722527980804, + 0.010261030867695808, + 0.013846024870872498, + 0.006647591013461351, + 0.011095619760453701, + 0.010662579908967018, + 0.008736619725823402, + 0.020189844071865082, + 0.02264409512281418, + 0.0295677799731493, + 0.019305182620882988, + 0.02066846564412117, + 0.019935419782996178, + 0.02173583209514618, + 0.031136643141508102, + 0.05338585749268532, + 0.03191183879971504, + 0.008825089782476425, + 0.01131231989711523, + 0.010821548290550709, + 0.009176668711006641, + 0.009418701753020287, + 0.014636123552918434, + 0.009032863192260265, + 0.012783819809556007, + 0.013651899062097073, + 0.015035945922136307, + 0.02038286253809929, + 0.023866353556513786, + 0.015528078190982342, + 0.026462728157639503, + 0.017923446372151375, + 0.02287001721560955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.05494515597820282 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, deliberate left-to-right pan, initiated by a brilliant, anamorphic lens flare that washes the frame in streaks of warm gold and ethereal blue light. As the flash gracefully sweeps across the lens, it reveals a low-angle shot of a truly enormous python, calmly coiled around a sun-drenched, mossy boulder in a humid jungle clearing. The snake's massive body is wound into a tight, intricate spiral, its scales a captivating iridescent mosaic of olive green and rich earthy brown, forming an intricate diamond pattern that shimmers in the intense sunlight. Each individual scale is so distinct that its texture and the way it catches the light are visible. The creature is in a state of deep rest; its head is tucked into its coil, the only perceptible movement being the slow, rhythmic expansion and contraction of its powerful abdomen as it breathes. The environment is lush and vibrant, with golden afternoon sunlight penetrating the dense jungle canopy above, dotting the forest floor and the snake's shimmering back with spots. The background is a soft-focus tapestry of deep green ferns, hanging vines, and the dark, fertile soil of the jungle. The overall atmosphere is one of profound tranquility and majestic power, capturing a quiet, awe-inspiring moment of nature in the warm, hazy glow of the sun.", + "seed": 31250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31250_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05328002125024796, + "vlm_logprob": -0.05328002125024796, + "vlm_yes_no_margin": 3.374999976158142, + "vlm_token_credit": -0.05328002125024796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06881190836429596 + ], + "opposite_logprob_tokens": [ + -2.8188118934631348 + ], + "logprob_reward": -0.06881190836429596, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06881190836429596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22977736592292786, + 0.10746310651302338, + 0.08836289495229721, + 0.10239043086767197, + 0.05206180363893509, + 0.0685073584318161, + 0.0388009250164032, + 0.05145598575472832, + 0.17647288739681244, + 0.08470729738473892 + ], + "credit_map": [ + 0.022171637043356895, + 0.01651754043996334, + 0.029723992571234703, + 0.02977105788886547, + 0.032103732228279114, + 0.03972732275724411, + 0.025256507098674774, + 0.03450554609298706, + 0.007992769591510296, + 0.013321547769010067, + 0.010143650695681572, + 0.019704030826687813, + 0.017284942790865898, + 0.014948626048862934, + 0.013914644718170166, + 0.010152891278266907, + 0.008019817993044853, + 0.009517987258732319, + 0.013361027464270592, + 0.014859020709991455, + 0.010634969919919968, + 0.011749434284865856, + 0.011774484999477863, + 0.00844613928347826, + 0.01030079647898674, + 0.011567256413400173, + 0.008416969329118729, + 0.012131287716329098, + 0.013780318200588226, + 0.013252489268779755, + 0.022106509655714035, + 0.010834794491529465, + 0.0056734103709459305, + 0.00625066040083766, + 0.006123215425759554, + 0.00665285112336278, + 0.006820195820182562, + 0.006468975450843573, + 0.0089704105630517, + 0.005102079827338457, + 0.0065689426846802235, + 0.0075042955577373505, + 0.009154264815151691, + 0.009306478314101696, + 0.008472736924886703, + 0.009933553636074066, + 0.007992078550159931, + 0.009575007483363152, + 0.0031955353915691376, + 0.004857675172388554, + 0.005505655892193317, + 0.0060216463170945644, + 0.00513208145275712, + 0.0050547984428703785, + 0.004553193226456642, + 0.004480335861444473, + 0.005610266700387001, + 0.00598853500559926, + 0.009500272572040558, + 0.005767051130533218, + 0.004429230932146311, + 0.007650760933756828, + 0.005626960191875696, + 0.006882902700453997, + 0.013539540581405163, + 0.013464837335050106, + 0.04026920720934868, + 0.019429702311754227, + 0.03275323659181595, + 0.023383328691124916, + 0.018633507192134857, + 0.014999516308307648, + 0.0071655274368822575, + 0.007714873179793358, + 0.011704103089869022, + 0.006694034207612276, + 0.011438380926847458, + 0.02144196815788746, + 0.006648665759712458, + 0.011899732053279877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.06881190836429596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869815334677696 + ], + "opposite_logprob_tokens": [ + -3.0486981868743896 + ], + "logprob_reward": -0.04869815334677696, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869815334677696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18267805874347687, + 0.16421198844909668, + 0.1825992912054062, + 0.09216950088739395, + 0.08351630717515945, + 0.03539266437292099, + 0.017058461904525757, + 0.0306132510304451, + 0.18066860735416412, + 0.03109186515212059 + ], + "credit_map": [ + 0.01777356117963791, + 0.015109003521502018, + 0.01765652559697628, + 0.028969336301088333, + 0.03435298055410385, + 0.02347717434167862, + 0.021108707413077354, + 0.024230780079960823, + 0.01053012628108263, + 0.013528494164347649, + 0.016977068036794662, + 0.041407838463783264, + 0.018646547570824623, + 0.019751794636249542, + 0.029619324952363968, + 0.01375079620629549, + 0.016785262152552605, + 0.02481728047132492, + 0.02003367803990841, + 0.028227319940924644, + 0.024157840758562088, + 0.031720686703920364, + 0.02136985957622528, + 0.0154873700812459, + 0.008163981139659882, + 0.016074474900960922, + 0.01160429883748293, + 0.014625991694629192, + 0.013124935328960419, + 0.011268670670688152, + 0.009999660775065422, + 0.0073074777610599995, + 0.008816092275083065, + 0.010646636597812176, + 0.01059658918529749, + 0.01176952850073576, + 0.012292098253965378, + 0.011716597713530064, + 0.01024730782955885, + 0.00743146101012826, + 0.00464342487975955, + 0.004404277075082064, + 0.0038887467235326767, + 0.005175052210688591, + 0.0040717963129282, + 0.005499341990798712, + 0.00395883247256279, + 0.0037511952687054873, + 0.0021120200399309397, + 0.0034031460527330637, + 0.002452478278428316, + 0.002225917298346758, + 0.001282486249692738, + 0.0024955093394964933, + 0.0016694128280505538, + 0.0014174918178468943, + 0.0033825826831161976, + 0.004368252586573362, + 0.004026505630463362, + 0.003994353115558624, + 0.003797894110903144, + 0.003671466140076518, + 0.003026291262358427, + 0.004345905967056751, + 0.01261816918849945, + 0.011180211789906025, + 0.05864952877163887, + 0.01191653124988079, + 0.008737755939364433, + 0.057550568133592606, + 0.009728698991239071, + 0.010287131182849407, + 0.005483108572661877, + 0.003910150844603777, + 0.004489274229854345, + 0.003688863478600979, + 0.003500433173030615, + 0.004219100344926119, + 0.0026877475902438164, + 0.0031131848227232695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.04869815334677696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632143795490265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632143795490265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0857757031917572, + 0.21882669627666473, + 0.144296795129776, + 0.06598321348428726, + 0.1055217832326889, + 0.04587104544043541, + 0.023962782695889473, + 0.08542261272668839, + 0.17236700654029846, + 0.05197235569357872 + ], + "credit_map": [ + 0.007366817910224199, + 0.007166756317019463, + 0.01107408944517374, + 0.018053339794278145, + 0.00865297019481659, + 0.014834587462246418, + 0.01059993077069521, + 0.008027219213545322, + 0.015193569473922253, + 0.023521360009908676, + 0.020326143130660057, + 0.04310382157564163, + 0.024578768759965897, + 0.03300522267818451, + 0.04234343022108078, + 0.016754385083913803, + 0.014425893314182758, + 0.01540193147957325, + 0.020095625892281532, + 0.023036271333694458, + 0.020013907924294472, + 0.021460840478539467, + 0.016805076971650124, + 0.01305724959820509, + 0.007752133067697287, + 0.011993772350251675, + 0.0074513754807412624, + 0.008933157660067081, + 0.009180841967463493, + 0.007868987508118153, + 0.007458637934178114, + 0.00534430705010891, + 0.01135067455470562, + 0.013339378871023655, + 0.01268032006919384, + 0.012141568586230278, + 0.016258856281638145, + 0.013423957861959934, + 0.01599264331161976, + 0.01033438928425312, + 0.00656146788969636, + 0.006092796567827463, + 0.004568050149828196, + 0.004484992008656263, + 0.004110796842724085, + 0.006413841154426336, + 0.007674264721572399, + 0.0059648361057043076, + 0.0029174922965466976, + 0.002991060959175229, + 0.003998745232820511, + 0.002720597665756941, + 0.002041771076619625, + 0.003923544194549322, + 0.0028419778682291508, + 0.0025275927037000656, + 0.009531543590128422, + 0.01423379685729742, + 0.01417362131178379, + 0.01031628530472517, + 0.006924889981746674, + 0.013434123247861862, + 0.008286257274448872, + 0.00852209609001875, + 0.011333751492202282, + 0.012255099602043629, + 0.04845167696475983, + 0.013639491982758045, + 0.020745599642395973, + 0.041649118065834045, + 0.012992379255592823, + 0.011299886740744114, + 0.005599403753876686, + 0.005911263171583414, + 0.00757875619456172, + 0.006801495794206858, + 0.0075449165888130665, + 0.0072270892560482025, + 0.005751385819166899, + 0.005558045115321875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200706034898758 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200706034898758, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200706034898758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14452551305294037, + 0.22197845578193665, + 0.09352302551269531, + 0.048002008348703384, + 0.04052480310201645, + 0.0791637971997261, + 0.022773969918489456, + 0.04279641434550285, + 0.23368486762046814, + 0.07302723079919815 + ], + "credit_map": [ + 0.011418061330914497, + 0.012395757250487804, + 0.0187566839158535, + 0.02115635760128498, + 0.019664790481328964, + 0.02885405719280243, + 0.014804885722696781, + 0.017474913969635963, + 0.013833646662533283, + 0.022642677649855614, + 0.017472870647907257, + 0.0395570807158947, + 0.03472830355167389, + 0.029403559863567352, + 0.04034015163779259, + 0.02400016225874424, + 0.007184592075645924, + 0.010803119279444218, + 0.013471182435750961, + 0.012879663147032261, + 0.012259051203727722, + 0.014153167605400085, + 0.014481096528470516, + 0.008291155099868774, + 0.004913961514830589, + 0.006735015660524368, + 0.005002220626920462, + 0.0064733210019767284, + 0.007450254168361425, + 0.005190001334995031, + 0.007299588527530432, + 0.0049376473762094975, + 0.004105915781110525, + 0.0046740947291255, + 0.005698465742170811, + 0.005141976289451122, + 0.005252521019428968, + 0.005038364790380001, + 0.006140179932117462, + 0.004473287146538496, + 0.007597474846988916, + 0.007525548804551363, + 0.00880709383636713, + 0.009894910268485546, + 0.006527696270495653, + 0.012907738797366619, + 0.011510821059346199, + 0.014392515644431114, + 0.002421750919893384, + 0.0029480778612196445, + 0.0028032376430928707, + 0.004054495599120855, + 0.00201810454018414, + 0.0031289192847907543, + 0.0024986632633954287, + 0.0029007208067923784, + 0.004237575922161341, + 0.0053563411347568035, + 0.008237666450440884, + 0.0048833428882062435, + 0.0042626927606761456, + 0.005657924804836512, + 0.004365015309303999, + 0.005795856472104788, + 0.014912398532032967, + 0.012959997169673443, + 0.07593410462141037, + 0.015429570339620113, + 0.013719093054533005, + 0.07553182542324066, + 0.013876542448997498, + 0.011321335099637508, + 0.011065630242228508, + 0.008129719644784927, + 0.008870422840118408, + 0.00717859948053956, + 0.013156522065401077, + 0.010589703917503357, + 0.0068569304421544075, + 0.007179697044193745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014200706034898758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705719470977783 + ], + "opposite_logprob_tokens": [ + -2.1270570755004883 + ], + "logprob_reward": -0.12705719470977783, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705719470977783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12863962352275848, + 0.2591397166252136, + 0.22927433252334595, + 0.045210063457489014, + 0.0625913143157959, + 0.04524583742022514, + 0.03168907389044762, + 0.06815855950117111, + 0.09515144675970078, + 0.03490002825856209 + ], + "credit_map": [ + 0.012309708632528782, + 0.011340869590640068, + 0.0167025588452816, + 0.01629476249217987, + 0.014953715726733208, + 0.01620689406991005, + 0.02445348910987377, + 0.016377612948417664, + 0.017866304144263268, + 0.020899932831525803, + 0.01974424533545971, + 0.04220440611243248, + 0.05115121975541115, + 0.033746637403964996, + 0.05013449117541313, + 0.023392481729388237, + 0.016859250143170357, + 0.024389708414673805, + 0.019088322296738625, + 0.02363905869424343, + 0.040944285690784454, + 0.03255762532353401, + 0.052172448486089706, + 0.01962362788617611, + 0.006093438249081373, + 0.00476784398779273, + 0.00422434089705348, + 0.006166009232401848, + 0.007701602298766375, + 0.005269452929496765, + 0.006017301697283983, + 0.004970069508999586, + 0.006391037721186876, + 0.008181299082934856, + 0.008110720664262772, + 0.008726861327886581, + 0.008940369822084904, + 0.00824105553328991, + 0.00870476197451353, + 0.005295204930007458, + 0.005546858534216881, + 0.005784465931355953, + 0.005044565070420504, + 0.0058134207502007484, + 0.0040164729580283165, + 0.005898207426071167, + 0.007151317782700062, + 0.005990525241941214, + 0.0028596692718565464, + 0.004204847849905491, + 0.004352219868451357, + 0.005568327847868204, + 0.003110910300165415, + 0.003933206666260958, + 0.00381586910225451, + 0.003844022750854492, + 0.005806686822324991, + 0.007012967020273209, + 0.012347177602350712, + 0.00825416948646307, + 0.0069272154942154884, + 0.011439729481935501, + 0.008380354382097721, + 0.00799025408923626, + 0.005193892866373062, + 0.008003413677215576, + 0.015937989577651024, + 0.012229766696691513, + 0.016204670071601868, + 0.015055137686431408, + 0.012411318719387054, + 0.010115256533026695, + 0.005354140419512987, + 0.0044364770874381065, + 0.004374371375888586, + 0.0036350360605865717, + 0.005526395980268717, + 0.003752795048058033, + 0.004646382760256529, + 0.0031744285952299833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.12705719470977783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, deliberate left-to-right pan, initiated by a brilliant, anamorphic lens flare that washes the frame in streaks of warm gold and ethereal blue light. As the flash gracefully sweeps across the lens, it reveals a low-angle shot of a truly enormous python, calmly coiled around a sun-drenched, mossy boulder in a humid jungle clearing. The snake's massive body is wound into a tight, intricate spiral, its scales a captivating iridescent mosaic of olive green and rich earthy brown, forming an intricate diamond pattern that shimmers in the intense sunlight. Each individual scale is so distinct that its texture and the way it catches the light are visible. The creature is in a state of deep rest; its head is tucked into its coil, the only perceptible movement being the slow, rhythmic expansion and contraction of its powerful abdomen as it breathes. The environment is lush and vibrant, with golden afternoon sunlight penetrating the dense jungle canopy above, dotting the forest floor and the snake's shimmering back with spots. The background is a soft-focus tapestry of deep green ferns, hanging vines, and the dark, fertile soil of the jungle. The overall atmosphere is one of profound tranquility and majestic power, capturing a quiet, awe-inspiring moment of nature in the warm, hazy glow of the sun.", + "seed": 31251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_2/index_5961_seed_31251_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052372701559215785, + "vlm_logprob": -0.052372701559215785, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.052372701559215785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11960934847593307 + ], + "opposite_logprob_tokens": [ + -2.2446093559265137 + ], + "logprob_reward": -0.11960934847593307, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11960934847593307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13458700478076935, + 0.15189972519874573, + 0.07353989034891129, + 0.08802072703838348, + 0.062164705246686935, + 0.16719050705432892, + 0.0832594782114029, + 0.04136066883802414, + 0.11354578286409378, + 0.08443157374858856 + ], + "credit_map": [ + 0.019236570224165916, + 0.014291482046246529, + 0.019116301089525223, + 0.01221495307981968, + 0.012094256468117237, + 0.019644299522042274, + 0.018431005999445915, + 0.019558135420084, + 0.013499265536665916, + 0.016784023493528366, + 0.022603372111916542, + 0.023214105516672134, + 0.020745478570461273, + 0.016828013584017754, + 0.021392665803432465, + 0.016832806169986725, + 0.00615218048915267, + 0.006401005666702986, + 0.008304302580654621, + 0.011605928651988506, + 0.008995503187179565, + 0.006361773237586021, + 0.015347637236118317, + 0.010371561162173748, + 0.006506173871457577, + 0.006523579824715853, + 0.00705783162266016, + 0.012916656211018562, + 0.0059539927169680595, + 0.006841856520622969, + 0.028757063671946526, + 0.013463574461638927, + 0.006231686100363731, + 0.006401970051229, + 0.00563473766669631, + 0.006898597348481417, + 0.009817096404731274, + 0.010445792227983475, + 0.008235090412199497, + 0.00849973689764738, + 0.006831004749983549, + 0.007743192370980978, + 0.012924056500196457, + 0.008388555608689785, + 0.005687346216291189, + 0.0071982089430093765, + 0.05686850845813751, + 0.06154964491724968, + 0.006955449003726244, + 0.008454586379230022, + 0.017029589042067528, + 0.005003747530281544, + 0.005886414088308811, + 0.009581727907061577, + 0.011442405171692371, + 0.01890556886792183, + 0.0032011407893151045, + 0.0029840231873095036, + 0.007783844601362944, + 0.0038001651410013437, + 0.003174400422722101, + 0.007933653891086578, + 0.006265175063163042, + 0.00621826434507966, + 0.008445865474641323, + 0.011164296418428421, + 0.021582163870334625, + 0.02997901663184166, + 0.00623279670253396, + 0.008640879765152931, + 0.01109408214688301, + 0.01640668697655201, + 0.010837700217962265, + 0.011498970910906792, + 0.01165963988751173, + 0.016579020768404007, + 0.005642317235469818, + 0.005016148556023836, + 0.008705959655344486, + 0.014491814188659191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.11960934847593307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903856784105301 + ], + "opposite_logprob_tokens": [ + -2.579038619995117 + ], + "logprob_reward": -0.07903856784105301, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903856784105301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2112775593996048, + 0.10969965904951096, + 0.05140494182705879, + 0.049424052238464355, + 0.07396075129508972, + 0.08510202914476395, + 0.06577570736408234, + 0.08332698792219162, + 0.06468842178583145, + 0.20533984899520874 + ], + "credit_map": [ + 0.025956707075238228, + 0.028461836278438568, + 0.04019938036799431, + 0.024660468101501465, + 0.01873067021369934, + 0.025156185030937195, + 0.02201465144753456, + 0.026097675785422325, + 0.011110659688711166, + 0.017086954787373543, + 0.017225395888090134, + 0.01581384800374508, + 0.012046387419104576, + 0.008379065431654453, + 0.0143242496997118, + 0.0137131093069911, + 0.006676117889583111, + 0.006802585907280445, + 0.0076193539425730705, + 0.007832103408873081, + 0.005429822485893965, + 0.004007741808891296, + 0.0061199781484901905, + 0.006917241495102644, + 0.0054848408326506615, + 0.004748492501676083, + 0.004437018185853958, + 0.006963189225643873, + 0.005247126799076796, + 0.004782141651958227, + 0.011432338505983353, + 0.006328906398266554, + 0.007964336313307285, + 0.00897762831300497, + 0.00879785418510437, + 0.01046043448150158, + 0.008311780169606209, + 0.006008779164403677, + 0.01137653086334467, + 0.01206341851502657, + 0.004676640033721924, + 0.00732127670198679, + 0.00887091550976038, + 0.007569488603621721, + 0.005424006842076778, + 0.011549643240869045, + 0.0199421476572752, + 0.01974792405962944, + 0.0038915444165468216, + 0.009937359020113945, + 0.01611783355474472, + 0.004698074422776699, + 0.0038091621827334166, + 0.005121408496052027, + 0.006687523797154427, + 0.015512799844145775, + 0.005176987498998642, + 0.008119693025946617, + 0.016852475702762604, + 0.008848336525261402, + 0.006996234878897667, + 0.012430462054908276, + 0.009437772445380688, + 0.015465031377971172, + 0.005282067693769932, + 0.006833742372691631, + 0.01763138920068741, + 0.01036493293941021, + 0.003404264571145177, + 0.004598330240696669, + 0.00551268644630909, + 0.01106101181358099, + 0.020003529265522957, + 0.024535953998565674, + 0.029216980561614037, + 0.05152909830212593, + 0.01379816047847271, + 0.012341818772256374, + 0.021380098536610603, + 0.03253421559929848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07903856784105301 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952369887381792 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952369887381792, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952369887381792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13302195072174072, + 0.21777532994747162, + 0.06769780069589615, + 0.09181337058544159, + 0.047026049345731735, + 0.13166286051273346, + 0.10562596470117569, + 0.045922037214040756, + 0.09566711634397507, + 0.0637875348329544 + ], + "credit_map": [ + 0.012333030812442303, + 0.015545959584414959, + 0.022476738318800926, + 0.014495845884084702, + 0.009790107607841492, + 0.023457098752260208, + 0.017546700313687325, + 0.017376456409692764, + 0.019032908603549004, + 0.03040774166584015, + 0.028643691912293434, + 0.026027806103229523, + 0.03729686513543129, + 0.015017767436802387, + 0.021937154233455658, + 0.03941139206290245, + 0.005720376502722502, + 0.005365496501326561, + 0.0068848324008286, + 0.010995150543749332, + 0.00636676512658596, + 0.007502603344619274, + 0.014434661716222763, + 0.010427913628518581, + 0.006465620826929808, + 0.01000536885112524, + 0.01132060494273901, + 0.014267642982304096, + 0.0084473742172122, + 0.008350647054612637, + 0.02114088088274002, + 0.011815233156085014, + 0.005893448367714882, + 0.006710781250149012, + 0.005191875156015158, + 0.004440721590071917, + 0.003919313196092844, + 0.0042568682692945, + 0.007914938032627106, + 0.008698100224137306, + 0.0063214488327503204, + 0.00636719586327672, + 0.013498639687895775, + 0.008130763657391071, + 0.005932466126978397, + 0.008480452001094818, + 0.04138360917568207, + 0.041548293083906174, + 0.007807567249983549, + 0.018320605158805847, + 0.016730351373553276, + 0.010325347073376179, + 0.008822114206850529, + 0.010717762634158134, + 0.014843517914414406, + 0.018058698624372482, + 0.004063992295414209, + 0.005171500146389008, + 0.009681455790996552, + 0.0045621502213180065, + 0.0035466179251670837, + 0.003935015760362148, + 0.006854151841253042, + 0.00810715276747942, + 0.007195590063929558, + 0.008836253546178341, + 0.022095758467912674, + 0.022659026086330414, + 0.004560172092169523, + 0.006617653649300337, + 0.009286796674132347, + 0.014415859244763851, + 0.005918234586715698, + 0.007706950418651104, + 0.009391465224325657, + 0.013048981316387653, + 0.0033562227617949247, + 0.005108935292810202, + 0.0069327387027442455, + 0.012324006296694279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005952369887381792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201764017343521 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201764017343521, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201764017343521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35235926508903503, + 0.24404089152812958, + 0.06764602661132812, + 0.04848258197307587, + 0.033103469759225845, + 0.07171743363142014, + 0.027236882597208023, + 0.06042330339550972, + 0.03717268258333206, + 0.05781743675470352 + ], + "credit_map": [ + 0.03662722557783127, + 0.04272862896323204, + 0.06831007450819016, + 0.04807751253247261, + 0.033707816153764725, + 0.04106558859348297, + 0.036562755703926086, + 0.045279648154973984, + 0.011961854062974453, + 0.023030877113342285, + 0.03363387659192085, + 0.04493134468793869, + 0.033200617879629135, + 0.02036181464791298, + 0.02601575292646885, + 0.050904758274555206, + 0.006738576572388411, + 0.007284617982804775, + 0.0077660265378654, + 0.011227820999920368, + 0.007072994951158762, + 0.0066070822067558765, + 0.01198998186737299, + 0.008958927355706692, + 0.0040370202623307705, + 0.004086943808943033, + 0.006456513423472643, + 0.0076700542122125626, + 0.004073362331837416, + 0.004079298581928015, + 0.010723557323217392, + 0.007355833891779184, + 0.003042580559849739, + 0.0034890056122094393, + 0.003141737310215831, + 0.0034129612613469362, + 0.002675443422049284, + 0.002966729924082756, + 0.006957631558179855, + 0.007417379412800074, + 0.0047173029743134975, + 0.004383210092782974, + 0.008074273355305195, + 0.0043978337198495865, + 0.006340923719108105, + 0.017177639529109, + 0.015447335317730904, + 0.01117892749607563, + 0.0019892859272658825, + 0.0035685391630977392, + 0.004630422219634056, + 0.002920794300734997, + 0.0021034886594861746, + 0.002470949897542596, + 0.004002036992460489, + 0.005551367998123169, + 0.005003954749554396, + 0.0064528021030128, + 0.006740746088325977, + 0.007730347104370594, + 0.006352548487484455, + 0.011197462677955627, + 0.00752685684710741, + 0.009418591856956482, + 0.004152350593358278, + 0.003205401124432683, + 0.006962964776903391, + 0.005249378737062216, + 0.0033197521697729826, + 0.003992269281297922, + 0.004444354213774204, + 0.005846214015036821, + 0.005689099431037903, + 0.005090487189590931, + 0.007656524423509836, + 0.010498409159481525, + 0.00571333896368742, + 0.00371897267177701, + 0.006665676832199097, + 0.012784933671355247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014201764017343521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a slow, deliberate left-to-right pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by an anamorphic lens flare with streaks of warm gold and ethereal blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a low-angle shot of an enormous python coiled around a sun-drenched, mossy boulder visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the python's scales display an iridescent mosaic of olive green and rich earthy brown in a diamond pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the snake's head tucked into its coil with only the slow expansion and contraction of its abdomen visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306145757436752 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.04306145757436752, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306145757436752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2527334988117218, + 0.10622334480285645, + 0.05976686626672745, + 0.05373009666800499, + 0.030539672821760178, + 0.08250030875205994, + 0.05860696732997894, + 0.05339255556464195, + 0.25232547521591187, + 0.05018122121691704 + ], + "credit_map": [ + 0.026687700301408768, + 0.033463042229413986, + 0.04560062661767006, + 0.03069993108510971, + 0.02189556695520878, + 0.03555568307638168, + 0.028628291562199593, + 0.03020264394581318, + 0.009047734551131725, + 0.00991582777351141, + 0.01327462401241064, + 0.014817699790000916, + 0.015276556834578514, + 0.015890398994088173, + 0.011135896667838097, + 0.016864599660038948, + 0.008251859806478024, + 0.007115352898836136, + 0.007062718737870455, + 0.009120414964854717, + 0.005685219541192055, + 0.005020895507186651, + 0.010208884254097939, + 0.007301521487534046, + 0.004779614508152008, + 0.004282904788851738, + 0.004981689620763063, + 0.008241357281804085, + 0.004215457011014223, + 0.004608811344951391, + 0.015561346895992756, + 0.007058915216475725, + 0.002634957665577531, + 0.0033704203087836504, + 0.0031190926674753428, + 0.004244732204824686, + 0.0040228236466646194, + 0.00420918595045805, + 0.00437436206266284, + 0.004564099479466677, + 0.003933115396648645, + 0.005733380559831858, + 0.01068678218871355, + 0.008139302022755146, + 0.00672188401222229, + 0.014994427561759949, + 0.01988190785050392, + 0.012409511022269726, + 0.0029315375722944736, + 0.004631440155208111, + 0.01671825535595417, + 0.0039816973730921745, + 0.002689196029677987, + 0.0038560323882848024, + 0.007254618685692549, + 0.016544189304113388, + 0.004583931993693113, + 0.0050986106507480145, + 0.007429016288369894, + 0.007064315024763346, + 0.004782333504408598, + 0.007345636375248432, + 0.008785491809248924, + 0.008303219452500343, + 0.02277626283466816, + 0.02468547224998474, + 0.04908442869782448, + 0.05418853089213371, + 0.018432440236210823, + 0.018364671617746353, + 0.02814832516014576, + 0.03664533421397209, + 0.003907097969204187, + 0.004902489483356476, + 0.0070208231918513775, + 0.011664049699902534, + 0.004200262948870659, + 0.004287601914256811, + 0.005352008622139692, + 0.008846886456012726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04306145757436752 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_tensor([8288, 8288])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_tensor([8288, 8288])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..06d6469c18a35459b9ecbbb70f20487b930f2d69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_tensor([8288, 8288])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a dizzying, dynamic tracking Dutch angle shot, a young character in period costume gallops across an endless expanse of golden sand dunes. Her long, dark brown hair flutters violently in the wind, and she wears a faded blue-green coarse coat. Her hands grip the reins tightly, her face a mixture of determination and a hint of barely concealed panic, her eyes occasionally glancing behind her. Her chestnut horse's hooves fly, its muscles bulging, and each powerful kick sends a cloud of fine, mist-like sand swirling in its wake. Behind them, an unimaginably massive sand castle slowly shifts in a manner that defies the laws of physics. Its fantastical spires and spiral staircases, constructed of shifting sand, cast a vast, ever-shifting shadow across the sloping horizon, as if pursued by a living, breathing beast. The whole world is bathed in the warm golden sunlight of the afternoon, presenting the iconic texture of Ghibli animation - the colors are saturated and soft, the sky is a hand-painted washed blue, dotted with a few fluffy white clouds, and the air is filled with a warm and magical atmosphere, rendering this tense escape into a magnificent picture full of movement and fairy tale colors.", + "seed": 18644, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18644_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07808818439953029, + "vlm_logprob": -0.07808818439953029, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.07808818439953029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08486731350421906 + ], + "opposite_logprob_tokens": [ + -2.584867238998413 + ], + "logprob_reward": -0.08486731350421906, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08486731350421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1276615858078003, + 0.06272584199905396, + 0.2589860260486603, + 0.20626968145370483, + 0.06550681591033936, + 0.02686532400548458, + 0.04719176143407822, + 0.039258722215890884, + 0.08553651720285416, + 0.0799977108836174 + ], + "credit_map": [ + 0.01406090147793293, + 0.012392083182930946, + 0.018859265372157097, + 0.023540986701846123, + 0.014591989107429981, + 0.015868445858359337, + 0.01562880165874958, + 0.012719114311039448, + 0.006023571360856295, + 0.0066813393495976925, + 0.006428950931876898, + 0.008780086413025856, + 0.009114877320826054, + 0.0078042177483439445, + 0.008265330456197262, + 0.009627464227378368, + 0.012877756729722023, + 0.02094593457877636, + 0.0260781217366457, + 0.06345440447330475, + 0.023680616170167923, + 0.01965468004345894, + 0.06923702359199524, + 0.023057488724589348, + 0.02122604288160801, + 0.01870325580239296, + 0.01919582299888134, + 0.016015615314245224, + 0.021031713113188744, + 0.04886142164468765, + 0.02081141620874405, + 0.040424395352602005, + 0.0074149612337350845, + 0.0069953626953065395, + 0.006245107855647802, + 0.004911963827908039, + 0.010447779670357704, + 0.007811946328729391, + 0.006862421985715628, + 0.014817274175584316, + 0.003279165830463171, + 0.0027428811881691217, + 0.002426028484478593, + 0.0032724710181355476, + 0.0039665233343839645, + 0.0028610792942345142, + 0.003906128229573369, + 0.004411045927554369, + 0.004426999948918819, + 0.00504213385283947, + 0.005128475371748209, + 0.006311127450317144, + 0.005949280224740505, + 0.006697038654237986, + 0.007405650801956654, + 0.006231056991964579, + 0.005386585835367441, + 0.0035553514026105404, + 0.0037597259506583214, + 0.004000052344053984, + 0.007106295321136713, + 0.005308868363499641, + 0.0054054176434874535, + 0.004736427683383226, + 0.010153891518712044, + 0.00616143923252821, + 0.006007385905832052, + 0.011471306905150414, + 0.02266162633895874, + 0.008380044251680374, + 0.011135613545775414, + 0.00956521276384592, + 0.008602475747466087, + 0.003932081162929535, + 0.009922035969793797, + 0.012415104545652866, + 0.012266627512872219, + 0.01612250879406929, + 0.008511153981089592, + 0.008225720375776291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.08486731350421906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813480794429779 + ], + "opposite_logprob_tokens": [ + -1.4063481092453003 + ], + "logprob_reward": -0.2813480794429779, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813480794429779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21319270133972168, + 0.04889272525906563, + 0.10882958769798279, + 0.05215909704566002, + 0.18210965394973755, + 0.027063822373747826, + 0.049720682203769684, + 0.05283245071768761, + 0.055249329656362534, + 0.20994995534420013 + ], + "credit_map": [ + 0.027052471414208412, + 0.02420920319855213, + 0.029762085527181625, + 0.03835713118314743, + 0.023878535255789757, + 0.02503044903278351, + 0.02467305213212967, + 0.020229782909154892, + 0.009261058643460274, + 0.006538174580782652, + 0.006503677926957607, + 0.007194760255515575, + 0.005367150064557791, + 0.003935280721634626, + 0.006289929151535034, + 0.0038026953116059303, + 0.011694539338350296, + 0.01029249932616949, + 0.012294912710785866, + 0.01593860611319542, + 0.014415948651731014, + 0.013592765666544437, + 0.0180685855448246, + 0.012531728483736515, + 0.004251726903021336, + 0.005618671886622906, + 0.005524821113795042, + 0.004755561240017414, + 0.0074472976848483086, + 0.008786267600953579, + 0.0059419600293040276, + 0.009832789190113544, + 0.024810155853629112, + 0.01888355053961277, + 0.02780449204146862, + 0.020300135016441345, + 0.02483920380473137, + 0.019118377938866615, + 0.019544024020433426, + 0.026809701696038246, + 0.002688209991902113, + 0.002540334826335311, + 0.004279420245438814, + 0.005017873365432024, + 0.0038216044194996357, + 0.0023923730477690697, + 0.0035476135089993477, + 0.0027763936668634415, + 0.004940272308886051, + 0.00469181127846241, + 0.005208619404584169, + 0.01016786228865385, + 0.004001548979431391, + 0.005403358954936266, + 0.0097658047452569, + 0.005541403312236071, + 0.007445419207215309, + 0.005288857501000166, + 0.005027684383094311, + 0.007050056476145983, + 0.007652447558939457, + 0.006098367273807526, + 0.007538674399256706, + 0.006730944383889437, + 0.006461794022470713, + 0.005136558320373297, + 0.00529013155028224, + 0.006503663491457701, + 0.009504952467978, + 0.005937035661190748, + 0.01023804396390915, + 0.006177152972668409, + 0.039227116852998734, + 0.012081905268132687, + 0.020701156929135323, + 0.06568864732980728, + 0.01889069378376007, + 0.011149713769555092, + 0.02016923762857914, + 0.022041495889425278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.2813480794429779 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190113827586174 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018190113827586174, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018190113827586174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22903674840927124, + 0.11087583750486374, + 0.09619292616844177, + 0.10236337035894394, + 0.03049473837018013, + 0.028546512126922607, + 0.04831094294786453, + 0.048190437257289886, + 0.10222198814153671, + 0.20376648008823395 + ], + "credit_map": [ + 0.028114890679717064, + 0.02767493762075901, + 0.03810083866119385, + 0.03401484340429306, + 0.03893072158098221, + 0.023044362664222717, + 0.021457573398947716, + 0.01769859343767166, + 0.014725017361342907, + 0.0131654292345047, + 0.021140674129128456, + 0.01689554750919342, + 0.015634434297680855, + 0.0074769724160432816, + 0.0138143515214324, + 0.008023412898182869, + 0.009976422414183617, + 0.009332910180091858, + 0.013615361414849758, + 0.014570527710020542, + 0.011875798925757408, + 0.012213055044412613, + 0.015995625406503677, + 0.008613226003944874, + 0.02165965549647808, + 0.015388322994112968, + 0.013300344347953796, + 0.012626973912119865, + 0.012083032168447971, + 0.008263597264885902, + 0.01077253744006157, + 0.008268912322819233, + 0.0032384423539042473, + 0.0029317147564142942, + 0.003969276789575815, + 0.004168671555817127, + 0.004234983120113611, + 0.0023058762308210135, + 0.003993630409240723, + 0.005652143619954586, + 0.0037462369073182344, + 0.002615164266899228, + 0.004568467382341623, + 0.0036546869669109583, + 0.0031331644859164953, + 0.0028982728254050016, + 0.0033728647977113724, + 0.004557655658572912, + 0.0061379107646644115, + 0.006764203775674105, + 0.004938005935400724, + 0.00556308776140213, + 0.005809876602143049, + 0.006069752387702465, + 0.006934234872460365, + 0.00609386945143342, + 0.006341082509607077, + 0.0043669939041137695, + 0.005495594814419746, + 0.005880037322640419, + 0.009532014839351177, + 0.005223339423537254, + 0.006951020564883947, + 0.004400353413075209, + 0.01041510608047247, + 0.007334958296269178, + 0.011990857310593128, + 0.008631488308310509, + 0.02222249284386635, + 0.0112361591309309, + 0.016606412827968597, + 0.013784505426883698, + 0.03823859989643097, + 0.012568014673888683, + 0.016639087349176407, + 0.051636192947626114, + 0.027988675981760025, + 0.014476566575467587, + 0.023023037239909172, + 0.019196314737200737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.018190113827586174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019384894985705614 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019384894985705614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019384894985705614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16936545073986053, + 0.1463325321674347, + 0.07183873653411865, + 0.04984002560377121, + 0.17997677624225616, + 0.08014479279518127, + 0.03041665442287922, + 0.055314596742391586, + 0.05594860017299652, + 0.16082189977169037 + ], + "credit_map": [ + 0.015336507931351662, + 0.017049893736839294, + 0.022961869835853577, + 0.021822698414325714, + 0.020587081089615822, + 0.023435451090335846, + 0.02141708694398403, + 0.026754841208457947, + 0.016971774399280548, + 0.0164326261729002, + 0.03184390068054199, + 0.02194654941558838, + 0.02152552269399166, + 0.010982668958604336, + 0.017016511410474777, + 0.009612984955310822, + 0.008568544872105122, + 0.008483728393912315, + 0.011745055206120014, + 0.012691998854279518, + 0.008993655443191528, + 0.005704150535166264, + 0.009894242510199547, + 0.005757361650466919, + 0.006017771549522877, + 0.004978570621460676, + 0.006499658338725567, + 0.008328855037689209, + 0.0050965952686965466, + 0.00550181744620204, + 0.005852950271219015, + 0.0075638070702552795, + 0.024045253172516823, + 0.01711549423635006, + 0.03455541282892227, + 0.03209315240383148, + 0.018407074734568596, + 0.015011012554168701, + 0.020471492782235146, + 0.01827787235379219, + 0.01213037595152855, + 0.008634013123810291, + 0.018707673996686935, + 0.010808508843183517, + 0.012270587496459484, + 0.005679241381585598, + 0.007516751065850258, + 0.004397636279463768, + 0.004863488953560591, + 0.005152423866093159, + 0.0030598652083426714, + 0.005344804842025042, + 0.002038787119090557, + 0.0023342063650488853, + 0.0048002200201153755, + 0.0028228594455868006, + 0.0072058155201375484, + 0.007125986274331808, + 0.0059017157182097435, + 0.005363211501389742, + 0.011702163144946098, + 0.006609375588595867, + 0.006794310640543699, + 0.004612016025930643, + 0.0071893902495503426, + 0.005693096201866865, + 0.006085008382797241, + 0.007074493449181318, + 0.008449616841971874, + 0.0073736365884542465, + 0.007790662348270416, + 0.0062926942482590675, + 0.030185597017407417, + 0.020589469000697136, + 0.018001791089773178, + 0.04408201575279236, + 0.010421495884656906, + 0.00853426568210125, + 0.013542182743549347, + 0.015465067699551582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0019384894985705614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096925724297762 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004096925724297762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096925724297762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23878157138824463, + 0.06270071119070053, + 0.08710480481386185, + 0.17175963521003723, + 0.05936627462506294, + 0.032438915222883224, + 0.054424501955509186, + 0.03513303026556969, + 0.08791034668684006, + 0.17038021981716156 + ], + "credit_map": [ + 0.028371084481477737, + 0.028612522408366203, + 0.029748082160949707, + 0.02799881435930729, + 0.038508541882038116, + 0.030916163697838783, + 0.025385932996869087, + 0.029240421950817108, + 0.005977912340313196, + 0.006450068671256304, + 0.009393212385475636, + 0.009108209051191807, + 0.009611187502741814, + 0.006980524864047766, + 0.008502990938723087, + 0.006676608230918646, + 0.007960637100040913, + 0.008340275846421719, + 0.01608062908053398, + 0.014490574598312378, + 0.01043986901640892, + 0.006338862236589193, + 0.01503687258809805, + 0.008417077362537384, + 0.02841923013329506, + 0.02117333374917507, + 0.021803133189678192, + 0.01751742511987686, + 0.020025694742798805, + 0.025974169373512268, + 0.01710124872624874, + 0.019745394587516785, + 0.007044460624456406, + 0.006670503411442041, + 0.009707553312182426, + 0.00873580388724804, + 0.010368254967033863, + 0.004189756233245134, + 0.0072123208083212376, + 0.005437623243778944, + 0.0048095183447003365, + 0.0037364293821156025, + 0.0062263221479952335, + 0.004106811247766018, + 0.005398590117692947, + 0.002205821918323636, + 0.0034657942596822977, + 0.002489624312147498, + 0.00776422955095768, + 0.009424212388694286, + 0.0046674092300236225, + 0.006940883118659258, + 0.005255541764199734, + 0.004441970027983189, + 0.010093589313328266, + 0.005836663767695427, + 0.005371821112930775, + 0.0033832944463938475, + 0.004119716119021177, + 0.004620265681296587, + 0.005093260668218136, + 0.00423075258731842, + 0.005195428617298603, + 0.0031184933613985777, + 0.01015822496265173, + 0.007057979702949524, + 0.010824757628142834, + 0.00999804213643074, + 0.018327565863728523, + 0.007978253997862339, + 0.011228728108108044, + 0.0123367914929986, + 0.03478807210922241, + 0.009636025875806808, + 0.015211963094770908, + 0.0543346144258976, + 0.012673677876591682, + 0.008147502318024635, + 0.017249412834644318, + 0.018338939175009727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.004096925724297762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dizzying, dynamic tracking Dutch angle shot, a young character in period costume gallops across an endless expanse of golden sand dunes. Her long, dark brown hair flutters violently in the wind, and she wears a faded blue-green coarse coat. Her hands grip the reins tightly, her face a mixture of determination and a hint of barely concealed panic, her eyes occasionally glancing behind her. Her chestnut horse's hooves fly, its muscles bulging, and each powerful kick sends a cloud of fine, mist-like sand swirling in its wake. Behind them, an unimaginably massive sand castle slowly shifts in a manner that defies the laws of physics. Its fantastical spires and spiral staircases, constructed of shifting sand, cast a vast, ever-shifting shadow across the sloping horizon, as if pursued by a living, breathing beast. The whole world is bathed in the warm golden sunlight of the afternoon, presenting the iconic texture of Ghibli animation - the colors are saturated and soft, the sky is a hand-painted washed blue, dotted with a few fluffy white clouds, and the air is filled with a warm and magical atmosphere, rendering this tense escape into a magnificent picture full of movement and fairy tale colors.", + "seed": 18645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18645_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.045930960308760406, + "vlm_logprob": -0.045930960308760406, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.045930960308760406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09424368292093277 + ], + "opposite_logprob_tokens": [ + -2.4692437648773193 + ], + "logprob_reward": -0.09424368292093277, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09424368292093277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13346663117408752, + 0.2541518807411194, + 0.11018554121255875, + 0.05429571494460106, + 0.04908362776041031, + 0.10199807584285736, + 0.04714673385024071, + 0.0422709695994854, + 0.03326316922903061, + 0.17413760721683502 + ], + "credit_map": [ + 0.016036828979849815, + 0.01449909619987011, + 0.013089625164866447, + 0.011772663332521915, + 0.01580727845430374, + 0.01843687705695629, + 0.02800665609538555, + 0.015817610546946526, + 0.02188316360116005, + 0.02861718088388443, + 0.015530352480709553, + 0.017432818189263344, + 0.05183945968747139, + 0.02333657070994377, + 0.06595581769943237, + 0.029556546360254288, + 0.012049855664372444, + 0.010990441776812077, + 0.015564538538455963, + 0.01424293965101242, + 0.013464335352182388, + 0.0221728403121233, + 0.009647611528635025, + 0.012052989564836025, + 0.00633148942142725, + 0.006380841135978699, + 0.005901528988033533, + 0.004512368701398373, + 0.009420657530426979, + 0.005045335274189711, + 0.010749535635113716, + 0.005953959189355373, + 0.005355549044907093, + 0.004969262517988682, + 0.008287079632282257, + 0.006937849801033735, + 0.005425439216196537, + 0.004601959604769945, + 0.007579065393656492, + 0.00592742720618844, + 0.01103518158197403, + 0.007455870043486357, + 0.023061947897076607, + 0.01708526536822319, + 0.009304405190050602, + 0.00905014481395483, + 0.012522656470537186, + 0.012482613325119019, + 0.0053664580918848515, + 0.004487928934395313, + 0.007322228047996759, + 0.007310875691473484, + 0.006299453787505627, + 0.006405025254935026, + 0.005303028039634228, + 0.004651742056012154, + 0.004530320875346661, + 0.0038931667804718018, + 0.0035076590720564127, + 0.00495274318382144, + 0.008259105496108532, + 0.005073090549558401, + 0.006565572693943977, + 0.005489311646670103, + 0.0032837004400789738, + 0.002797719556838274, + 0.003137747524306178, + 0.0033546790946274996, + 0.005525936838239431, + 0.003463053610175848, + 0.005495901219546795, + 0.006204433739185333, + 0.016513755545020103, + 0.018102677538990974, + 0.026870133355259895, + 0.014367887750267982, + 0.030069755390286446, + 0.01870197430253029, + 0.02420458197593689, + 0.025306858122348785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.09424368292093277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033135116100311 + ], + "opposite_logprob_tokens": [ + -2.3503313064575195 + ], + "logprob_reward": -0.10033135116100311, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033135116100311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0747210681438446, + 0.1458626687526703, + 0.07215231657028198, + 0.03578188270330429, + 0.04056205973029137, + 0.037707120180130005, + 0.04496537521481514, + 0.04002799838781357, + 0.04138815775513649, + 0.46683135628700256 + ], + "credit_map": [ + 0.010440942831337452, + 0.00903275515884161, + 0.007816840894520283, + 0.008094764314591885, + 0.010085769928991795, + 0.005784759297966957, + 0.015105280093848705, + 0.008359954692423344, + 0.012408699840307236, + 0.01592613384127617, + 0.009686372242867947, + 0.009613456204533577, + 0.027116650715470314, + 0.013979348354041576, + 0.038832224905490875, + 0.018299777060747147, + 0.008267265744507313, + 0.00745217502117157, + 0.0119961928576231, + 0.010060527361929417, + 0.00836743600666523, + 0.011299279518425465, + 0.007639375980943441, + 0.007070062216371298, + 0.004773537162691355, + 0.003798056161031127, + 0.007616738323122263, + 0.005142034962773323, + 0.0027434455696493387, + 0.0029974880162626505, + 0.0044112964533269405, + 0.004299284890294075, + 0.005437502637505531, + 0.004124850500375032, + 0.006354286335408688, + 0.0051588257774710655, + 0.0062899538315832615, + 0.004885655362159014, + 0.004495284520089626, + 0.003815700765699148, + 0.004314390476793051, + 0.0029241295997053385, + 0.00922983419150114, + 0.006062743719667196, + 0.0035373088903725147, + 0.003431537887081504, + 0.003665866330265999, + 0.004541311413049698, + 0.004984874743968248, + 0.004223458003252745, + 0.004263587761670351, + 0.004960091318935156, + 0.006356744561344385, + 0.006808395031839609, + 0.00774426618590951, + 0.005623961333185434, + 0.004754433874040842, + 0.003630547085776925, + 0.004412636626511812, + 0.005483620800077915, + 0.006079583894461393, + 0.003593326546251774, + 0.006523661315441132, + 0.005550187546759844, + 0.003941787406802177, + 0.003610810963436961, + 0.00359336007386446, + 0.005238146521151066, + 0.005948249716311693, + 0.002678032498806715, + 0.009481116198003292, + 0.006896652281284332, + 0.03556044027209282, + 0.043161116540431976, + 0.05490492284297943, + 0.04445121809840202, + 0.06430094689130783, + 0.04927220195531845, + 0.10930071026086807, + 0.06587980687618256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.10033135116100311 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634979411959648 + ], + "opposite_logprob_tokens": [ + -3.6513497829437256 + ], + "logprob_reward": -0.02634979411959648, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634979411959648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21820604801177979, + 0.12604820728302002, + 0.059856116771698, + 0.07667671889066696, + 0.07254673540592194, + 0.04492830112576485, + 0.0657176673412323, + 0.07768721878528595, + 0.05643714964389801, + 0.2018958181142807 + ], + "credit_map": [ + 0.028084352612495422, + 0.025115443393588066, + 0.025975370779633522, + 0.025105509907007217, + 0.03405170515179634, + 0.02193266898393631, + 0.0339263454079628, + 0.02401464805006981, + 0.013166602700948715, + 0.014829431660473347, + 0.013686551712453365, + 0.013332056812942028, + 0.02325446344912052, + 0.012525919824838638, + 0.022087302058935165, + 0.013165878131985664, + 0.005680335219949484, + 0.007157635875046253, + 0.0058473520912230015, + 0.006489279214292765, + 0.01177285611629486, + 0.007541129365563393, + 0.009115740656852722, + 0.0062517947517335415, + 0.011398863978683949, + 0.009084234945476055, + 0.011890334077179432, + 0.010364185087382793, + 0.008058486506342888, + 0.007110943086445332, + 0.01101716049015522, + 0.007752514444291592, + 0.009799838997423649, + 0.007290289271622896, + 0.015266652218997478, + 0.01562589220702648, + 0.00573840131983161, + 0.004760051611810923, + 0.00623959070071578, + 0.007826016284525394, + 0.0061996616423130035, + 0.004547879099845886, + 0.0044427551329135895, + 0.004304933361709118, + 0.005870299879461527, + 0.00700257858261466, + 0.007503120228648186, + 0.00505707086995244, + 0.005852594971656799, + 0.005431805737316608, + 0.006903382483869791, + 0.009155451320111752, + 0.01101413182914257, + 0.009028648026287556, + 0.010756511241197586, + 0.007575141731649637, + 0.007654769811779261, + 0.006481164135038853, + 0.007685645017772913, + 0.009972949512302876, + 0.015548049472272396, + 0.008336999453604221, + 0.011924096383154392, + 0.010083544068038464, + 0.005429319571703672, + 0.0054145692847669125, + 0.005999721586704254, + 0.005956194829195738, + 0.007742896676063538, + 0.004657632205635309, + 0.012478094547986984, + 0.008758721873164177, + 0.01242965366691351, + 0.016825003549456596, + 0.02297598496079445, + 0.034602031111717224, + 0.024339260533452034, + 0.015677547082304955, + 0.04391447827219963, + 0.03113185055553913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02634979411959648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634353797882795 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634353797882795, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634353797882795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1311364471912384, + 0.22925585508346558, + 0.043761514127254486, + 0.059695690870285034, + 0.11343567073345184, + 0.045545849949121475, + 0.05460143834352493, + 0.06169433519244194, + 0.03303365781903267, + 0.22783948481082916 + ], + "credit_map": [ + 0.0162897240370512, + 0.013739625923335552, + 0.014055863954126835, + 0.013032673858106136, + 0.019166739657521248, + 0.023410513997077942, + 0.01853087544441223, + 0.012910435907542706, + 0.022140976041555405, + 0.02909316122531891, + 0.017770372331142426, + 0.01906193234026432, + 0.047472935169935226, + 0.019706349819898605, + 0.04938535392284393, + 0.024624763056635857, + 0.005164140835404396, + 0.0044504073448479176, + 0.00674830935895443, + 0.005650009494274855, + 0.005146103911101818, + 0.007822029292583466, + 0.004502888768911362, + 0.004277622327208519, + 0.007547030691057444, + 0.005770395044237375, + 0.010504195466637611, + 0.0091024711728096, + 0.005036945920437574, + 0.006550549529492855, + 0.007261860184371471, + 0.00792224146425724, + 0.013527424074709415, + 0.0100858174264431, + 0.021095285192131996, + 0.024145055562257767, + 0.006816249340772629, + 0.013485262170433998, + 0.011800382286310196, + 0.012480199337005615, + 0.007365517783910036, + 0.0049202656373381615, + 0.009370403364300728, + 0.006646655034273863, + 0.0038993139751255512, + 0.003774987068027258, + 0.0044123330153524876, + 0.005156372673809528, + 0.007444491609930992, + 0.006267012096941471, + 0.008116626180708408, + 0.009330205619335175, + 0.006129820831120014, + 0.00573147926479578, + 0.00566834257915616, + 0.005913460161536932, + 0.006926869973540306, + 0.005745215807110071, + 0.0075736697763204575, + 0.010797607712447643, + 0.008343156427145004, + 0.008719691075384617, + 0.006307448260486126, + 0.007280674763023853, + 0.0042583514004945755, + 0.003584379330277443, + 0.0039328234270215034, + 0.004191806074231863, + 0.003750795964151621, + 0.004132058005779982, + 0.005145637784153223, + 0.004037804435938597, + 0.030722271651029587, + 0.025435278192162514, + 0.03591904789209366, + 0.020863745361566544, + 0.022689085453748703, + 0.02620680257678032, + 0.03742605075240135, + 0.028577223420143127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.004634353797882795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17877544462680817, + 0.09492845833301544, + 0.04766954109072685, + 0.07335525006055832, + 0.10740590840578079, + 0.11042096465826035, + 0.07702929526567459, + 0.0981888622045517, + 0.03863212838768959, + 0.1735941469669342 + ], + "credit_map": [ + 0.025647297501564026, + 0.023629676550626755, + 0.02288816124200821, + 0.0194235946983099, + 0.022071994841098785, + 0.018736738711595535, + 0.02825940027832985, + 0.01811857521533966, + 0.010139189660549164, + 0.011547218076884747, + 0.0102002564817667, + 0.012845384888350964, + 0.016356732696294785, + 0.008647887967526913, + 0.015591729432344437, + 0.00960006657987833, + 0.005968746263533831, + 0.005337822251021862, + 0.009143704548478127, + 0.007616719231009483, + 0.004288733936846256, + 0.004795542452484369, + 0.005345859564840794, + 0.005172412842512131, + 0.0092916339635849, + 0.006629898212850094, + 0.01570107974112034, + 0.012053398415446281, + 0.004528423771262169, + 0.009539561346173286, + 0.007304088678210974, + 0.008307164534926414, + 0.01297073159366846, + 0.010643240995705128, + 0.02351524494588375, + 0.01676296629011631, + 0.012496919371187687, + 0.006443373393267393, + 0.013691766187548637, + 0.010881674475967884, + 0.014400537125766277, + 0.009460837580263615, + 0.023814234882593155, + 0.017000684514641762, + 0.007986362092196941, + 0.013079077005386353, + 0.011016037315130234, + 0.013663194142282009, + 0.011683516204357147, + 0.008502347394824028, + 0.012712178751826286, + 0.011713157407939434, + 0.009032795205712318, + 0.008213103748857975, + 0.008190823718905449, + 0.006981367245316505, + 0.01053039450198412, + 0.008744139224290848, + 0.009342399425804615, + 0.009207895025610924, + 0.01744057983160019, + 0.00875367596745491, + 0.01776101440191269, + 0.016408775001764297, + 0.004396616946905851, + 0.0032743855845183134, + 0.004546415992081165, + 0.0057279979810118675, + 0.005605657119303942, + 0.0037153244484215975, + 0.006592223420739174, + 0.004773504566401243, + 0.014170066453516483, + 0.015422863885760307, + 0.02354404516518116, + 0.01889723911881447, + 0.027409518137574196, + 0.02501554787158966, + 0.028465645387768745, + 0.020669231191277504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.004095619544386864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dizzying, dynamic tracking Dutch angle shot, a young character in period costume gallops across an endless expanse of golden sand dunes. Her long, dark brown hair flutters violently in the wind, and she wears a faded blue-green coarse coat. Her hands grip the reins tightly, her face a mixture of determination and a hint of barely concealed panic, her eyes occasionally glancing behind her. Her chestnut horse's hooves fly, its muscles bulging, and each powerful kick sends a cloud of fine, mist-like sand swirling in its wake. Behind them, an unimaginably massive sand castle slowly shifts in a manner that defies the laws of physics. Its fantastical spires and spiral staircases, constructed of shifting sand, cast a vast, ever-shifting shadow across the sloping horizon, as if pursued by a living, breathing beast. The whole world is bathed in the warm golden sunlight of the afternoon, presenting the iconic texture of Ghibli animation - the colors are saturated and soft, the sky is a hand-painted washed blue, dotted with a few fluffy white clouds, and the air is filled with a warm and magical atmosphere, rendering this tense escape into a magnificent picture full of movement and fairy tale colors.", + "seed": 18646, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18646_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06726859859190881, + "vlm_logprob": -0.06726859859190881, + "vlm_yes_no_margin": 3.574999976158142, + "vlm_token_credit": -0.06726859859190881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13270580768585205 + ], + "opposite_logprob_tokens": [ + -2.1327056884765625 + ], + "logprob_reward": -0.13270580768585205, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13270580768585205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1662287712097168, + 0.120612733066082, + 0.03682078421115875, + 0.16175657510757446, + 0.15503937005996704, + 0.047755561769008636, + 0.062034960836172104, + 0.04917525500059128, + 0.14561055600643158, + 0.05496535822749138 + ], + "credit_map": [ + 0.019367501139640808, + 0.016328811645507812, + 0.015245864167809486, + 0.013279814273118973, + 0.0422930046916008, + 0.017312172800302505, + 0.025103576481342316, + 0.017298033460974693, + 0.014564248733222485, + 0.010264587588608265, + 0.01690095104277134, + 0.021897578611969948, + 0.018952200189232826, + 0.012234663590788841, + 0.015125255100429058, + 0.01067324634641409, + 0.004079000558704138, + 0.0038848023395985365, + 0.006272675935178995, + 0.008970980532467365, + 0.004066077992320061, + 0.004322571214288473, + 0.0026113642379641533, + 0.0026133123319596052, + 0.03890987113118172, + 0.027202876284718513, + 0.0350228026509285, + 0.014262746088206768, + 0.009516425430774689, + 0.01770276203751564, + 0.008761363103985786, + 0.010377726517617702, + 0.01263335905969143, + 0.017541470006108284, + 0.0176673773676157, + 0.02165125496685505, + 0.03381185606122017, + 0.015315341763198376, + 0.016799623146653175, + 0.01961909420788288, + 0.00745828403159976, + 0.007845094427466393, + 0.006497013382613659, + 0.005930120591074228, + 0.010601853020489216, + 0.00420583039522171, + 0.002763057127594948, + 0.002454312052577734, + 0.005597539246082306, + 0.00613556569442153, + 0.009890745393931866, + 0.014073548838496208, + 0.0069181229919195175, + 0.007297806907445192, + 0.005534918513149023, + 0.006586709525436163, + 0.007350475061684847, + 0.006647814996540546, + 0.005829895846545696, + 0.006841916590929031, + 0.006961100269109011, + 0.003957834560424089, + 0.005081840790808201, + 0.006504378281533718, + 0.017333943396806717, + 0.01702781952917576, + 0.016825444996356964, + 0.017147142440080643, + 0.012111586518585682, + 0.018238598480820656, + 0.023503609001636505, + 0.023422420024871826, + 0.005535405594855547, + 0.0060063996352255344, + 0.007528743240982294, + 0.008408737368881702, + 0.005116576794534922, + 0.007249858696013689, + 0.005017733201384544, + 0.010101903229951859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.13270580768585205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603747010231018 + ], + "opposite_logprob_tokens": [ + -1.910374641418457 + ], + "logprob_reward": -0.1603747010231018, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603747010231018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11151868849992752, + 0.06998063623905182, + 0.05607713758945465, + 0.19000335037708282, + 0.05200103670358658, + 0.06134933605790138, + 0.2843610942363739, + 0.0706024095416069, + 0.07326124608516693, + 0.030845029279589653 + ], + "credit_map": [ + 0.012960216961801052, + 0.011681345291435719, + 0.01360698789358139, + 0.010853647254407406, + 0.025018824264407158, + 0.01267447043210268, + 0.014717631973326206, + 0.010005571879446507, + 0.00828964076936245, + 0.007103218697011471, + 0.010566997341811657, + 0.015537647530436516, + 0.010179363191127777, + 0.005972650367766619, + 0.005388048943132162, + 0.00694307079538703, + 0.008159846067428589, + 0.009361808188259602, + 0.009861799888312817, + 0.00901151541620493, + 0.005003875121474266, + 0.007000321056693792, + 0.003550978610292077, + 0.004127000458538532, + 0.06132768839597702, + 0.019511627033352852, + 0.053846392780542374, + 0.01264885626733303, + 0.007614503148943186, + 0.018764078617095947, + 0.00684808474034071, + 0.009442138485610485, + 0.006067189853638411, + 0.0054015954956412315, + 0.007253775838762522, + 0.00837758556008339, + 0.0058101266622543335, + 0.009445120580494404, + 0.00525980768725276, + 0.004385838285088539, + 0.007059980649501085, + 0.011236520484089851, + 0.007720931898802519, + 0.010920950211584568, + 0.01062107179313898, + 0.0056860740296542645, + 0.003860187716782093, + 0.004243625793606043, + 0.02797641046345234, + 0.03817308694124222, + 0.042689044028520584, + 0.0636921152472496, + 0.02003626897931099, + 0.047083768993616104, + 0.022061048075556755, + 0.022649375721812248, + 0.007903559133410454, + 0.007347135804593563, + 0.009454314596951008, + 0.01284603588283062, + 0.011017374694347382, + 0.006529228296130896, + 0.00853199977427721, + 0.0069727711379528046, + 0.00927400216460228, + 0.00881120003759861, + 0.011152682825922966, + 0.009635948576033115, + 0.008321376517415047, + 0.010080642066895962, + 0.007974750362336636, + 0.008010650984942913, + 0.0035784500651061535, + 0.003977981861680746, + 0.004021971486508846, + 0.004958262201398611, + 0.0038295085541903973, + 0.0036643780767917633, + 0.003208528971299529, + 0.003605953184887767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.1603747010231018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370017558336258 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370017558336258, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370017558336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15182967483997345, + 0.12161387503147125, + 0.03961707651615143, + 0.17297431826591492, + 0.06385564804077148, + 0.05627329275012016, + 0.21057221293449402, + 0.01867472566664219, + 0.13212645053863525, + 0.032462745904922485 + ], + "credit_map": [ + 0.018420062959194183, + 0.016533588990569115, + 0.013519604690372944, + 0.03163883090019226, + 0.02134702168405056, + 0.0145911555737257, + 0.020784681662917137, + 0.014994718134403229, + 0.015408223494887352, + 0.013064974918961525, + 0.019470559433102608, + 0.021406827494502068, + 0.021072933450341225, + 0.009837779216468334, + 0.011067424900829792, + 0.01028516050428152, + 0.004359472543001175, + 0.004321292042732239, + 0.005890918895602226, + 0.00846310518682003, + 0.005621916148811579, + 0.004163900390267372, + 0.003273800015449524, + 0.0035226696636527777, + 0.03131060674786568, + 0.03560614958405495, + 0.023072529584169388, + 0.021769700571894646, + 0.01379839051514864, + 0.02315397933125496, + 0.010867471806704998, + 0.013395493850111961, + 0.00593553576618433, + 0.007641117088496685, + 0.007226451765745878, + 0.0064314561896026134, + 0.01648123562335968, + 0.007253916934132576, + 0.007018317934125662, + 0.005867622792720795, + 0.0068167420104146, + 0.008564176969230175, + 0.005955977365374565, + 0.004747351631522179, + 0.01605241186916828, + 0.005624815821647644, + 0.004659250378608704, + 0.0038525681011378765, + 0.024531332775950432, + 0.02843710593879223, + 0.033104799687862396, + 0.045544251799583435, + 0.014645036309957504, + 0.03296291083097458, + 0.015052299946546555, + 0.016294479370117188, + 0.002528817392885685, + 0.002384078223258257, + 0.002610269235447049, + 0.0028955659363418818, + 0.002543120412155986, + 0.0021256550680845976, + 0.001862698351033032, + 0.0017245206981897354, + 0.017334850504994392, + 0.01650298573076725, + 0.019059734418988228, + 0.014833822846412659, + 0.015657322481274605, + 0.018917255103588104, + 0.01259719580411911, + 0.017223289236426353, + 0.004037519451230764, + 0.0028575786855071783, + 0.00445787888020277, + 0.003518529701977968, + 0.0024381615221500397, + 0.0050805481150746346, + 0.003453628160059452, + 0.00661889836192131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.03370017558336258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05976619943976402, + 0.08602554351091385, + 0.03210154175758362, + 0.22957062721252441, + 0.3554714024066925, + 0.023453665897250175, + 0.0412488617002964, + 0.02031872794032097, + 0.13552628457546234, + 0.016517138108611107 + ], + "credit_map": [ + 0.00700192479416728, + 0.005883151665329933, + 0.006207474507391453, + 0.008800688199698925, + 0.012308568693697453, + 0.005768828559666872, + 0.007056903094053268, + 0.006738661322742701, + 0.010219554416835308, + 0.018349559977650642, + 0.011686118319630623, + 0.010835409164428711, + 0.010732998140156269, + 0.009510653093457222, + 0.007056485395878553, + 0.007634765934199095, + 0.004329610615968704, + 0.005363642703741789, + 0.005804252345114946, + 0.0050829206593334675, + 0.003994672559201717, + 0.003115433966740966, + 0.0021026073954999447, + 0.002308401046320796, + 0.08300971984863281, + 0.026924213394522667, + 0.07244932651519775, + 0.013892272487282753, + 0.006436443421989679, + 0.011112814769148827, + 0.006332157179713249, + 0.009413671679794788, + 0.02512003667652607, + 0.037598881870508194, + 0.04393399506807327, + 0.04975900799036026, + 0.07262581586837769, + 0.05036427825689316, + 0.03532233089208603, + 0.040747057646512985, + 0.003098076907917857, + 0.004079548642039299, + 0.003207600675523281, + 0.0032528722658753395, + 0.0029980111867189407, + 0.0037502313498407602, + 0.0014363977825269103, + 0.0016309275524690747, + 0.004234843421727419, + 0.006105998530983925, + 0.0075712185353040695, + 0.00954535510390997, + 0.003273898037150502, + 0.004765400197356939, + 0.0028540000785142183, + 0.0028981438372284174, + 0.0022680023685097694, + 0.0024471916258335114, + 0.0030904661398380995, + 0.003279740223661065, + 0.002431359374895692, + 0.003521372564136982, + 0.001714840647764504, + 0.0015657529002055526, + 0.01893496699631214, + 0.016714122146368027, + 0.019749969244003296, + 0.016620375216007233, + 0.014642027206718922, + 0.019463200122117996, + 0.013150311075150967, + 0.016251303255558014, + 0.0024142079055309296, + 0.0022279901895672083, + 0.0020189080387353897, + 0.002234878484159708, + 0.0022922013886272907, + 0.0018770822789520025, + 0.0012340870453044772, + 0.0022177831269800663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.00361420470289886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948103964328766 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948103964328766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948103964328766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0670173242688179, + 0.031368523836135864, + 0.0272331815212965, + 0.3081131875514984, + 0.05485350266098976, + 0.03517141193151474, + 0.20365457236766815, + 0.020108317956328392, + 0.23516446352005005, + 0.01731550320982933 + ], + "credit_map": [ + 0.00726648885756731, + 0.007223396562039852, + 0.007923056371510029, + 0.0076116714626550674, + 0.014653841964900494, + 0.005852376576513052, + 0.008688907139003277, + 0.007797585800290108, + 0.0038897457998245955, + 0.002311291405931115, + 0.0038937737699598074, + 0.0034233503974974155, + 0.005765976849943399, + 0.003667069599032402, + 0.005059636197984219, + 0.003357677487656474, + 0.0035002322401851416, + 0.0020678038708865643, + 0.0048085786402225494, + 0.005833491683006287, + 0.002732102293521166, + 0.004156933166086674, + 0.001996654085814953, + 0.0021373843774199486, + 0.09478630870580673, + 0.04817045480012894, + 0.08395995944738388, + 0.02193578891456127, + 0.012413489632308483, + 0.021486101672053337, + 0.009575793519616127, + 0.015785301104187965, + 0.006096723489463329, + 0.005258757155388594, + 0.009375306777656078, + 0.00860255304723978, + 0.005341111216694117, + 0.009874556213617325, + 0.005230716895312071, + 0.00507377739995718, + 0.006354693789035082, + 0.003941784147173166, + 0.005059587769210339, + 0.004715592134743929, + 0.0040735723450779915, + 0.004907509777694941, + 0.00237103714607656, + 0.003747632959857583, + 0.021245427429676056, + 0.027809280902147293, + 0.03162868693470955, + 0.04474225267767906, + 0.014541659504175186, + 0.03166712075471878, + 0.015588265843689442, + 0.01643187738955021, + 0.0032224836759269238, + 0.002127883955836296, + 0.003194565186277032, + 0.0031288634054362774, + 0.001982045127078891, + 0.0022719530388712883, + 0.00222756271250546, + 0.0019529616693034768, + 0.02559596486389637, + 0.03082623891532421, + 0.03006766550242901, + 0.0257540475577116, + 0.03536548092961311, + 0.032273609191179276, + 0.026843557134270668, + 0.028437882661819458, + 0.0028994670137763023, + 0.0019043657230213284, + 0.002134930109605193, + 0.002153536770492792, + 0.0019325989997014403, + 0.002465620171278715, + 0.0013916913885623217, + 0.0024332941975444555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.005948103964328766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dizzying, dynamic tracking Dutch angle shot, a young character in period costume gallops across an endless expanse of golden sand dunes. Her long, dark brown hair flutters violently in the wind, and she wears a faded blue-green coarse coat. Her hands grip the reins tightly, her face a mixture of determination and a hint of barely concealed panic, her eyes occasionally glancing behind her. Her chestnut horse's hooves fly, its muscles bulging, and each powerful kick sends a cloud of fine, mist-like sand swirling in its wake. Behind them, an unimaginably massive sand castle slowly shifts in a manner that defies the laws of physics. Its fantastical spires and spiral staircases, constructed of shifting sand, cast a vast, ever-shifting shadow across the sloping horizon, as if pursued by a living, breathing beast. The whole world is bathed in the warm golden sunlight of the afternoon, presenting the iconic texture of Ghibli animation - the colors are saturated and soft, the sky is a hand-painted washed blue, dotted with a few fluffy white clouds, and the air is filled with a warm and magical atmosphere, rendering this tense escape into a magnificent picture full of movement and fairy tale colors.", + "seed": 18647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_3/index_8288_seed_18647_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08190402807667851, + "vlm_logprob": -0.08190402807667851, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.08190402807667851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10871231555938721 + ], + "opposite_logprob_tokens": [ + -2.3587121963500977 + ], + "logprob_reward": -0.10871231555938721, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10871231555938721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21122267842292786, + 0.09275350719690323, + 0.09863469004631042, + 0.0660422220826149, + 0.03290775790810585, + 0.04502837732434273, + 0.05747915059328079, + 0.21964643895626068, + 0.12318706512451172, + 0.05309811234474182 + ], + "credit_map": [ + 0.020570818334817886, + 0.019870471209287643, + 0.01964658312499523, + 0.0208240058273077, + 0.038969285786151886, + 0.02782572992146015, + 0.03642355278134346, + 0.027092227712273598, + 0.009135480970144272, + 0.005610795691609383, + 0.009989789687097073, + 0.007421639747917652, + 0.016192222014069557, + 0.010761159472167492, + 0.021022075787186623, + 0.012620346620678902, + 0.013505847193300724, + 0.011260372586548328, + 0.016565080732107162, + 0.01440848782658577, + 0.010431017726659775, + 0.010788765735924244, + 0.009971658699214458, + 0.01170345675200224, + 0.006620147731155157, + 0.00672632921487093, + 0.006122482009232044, + 0.008363899774849415, + 0.008736060932278633, + 0.005201841238886118, + 0.016373321413993835, + 0.007898140698671341, + 0.002619409002363682, + 0.002353868680074811, + 0.005934940185397863, + 0.0038906363770365715, + 0.002818820998072624, + 0.005496722646057606, + 0.004809810779988766, + 0.004983547143638134, + 0.007036518771201372, + 0.004807001445442438, + 0.00617985287681222, + 0.005789384711533785, + 0.00463991891592741, + 0.004742541816085577, + 0.005904267076402903, + 0.005928891245275736, + 0.004485918674618006, + 0.004947677254676819, + 0.006547883618623018, + 0.012449322268366814, + 0.007213185541331768, + 0.004895167425274849, + 0.009318884462118149, + 0.0076211076229810715, + 0.0168837308883667, + 0.026232540607452393, + 0.028105294331908226, + 0.0270113293081522, + 0.03538364917039871, + 0.024833736941218376, + 0.03705638274550438, + 0.024139773100614548, + 0.009404153563082218, + 0.01520905364304781, + 0.010546986013650894, + 0.013825326226651669, + 0.026199039071798325, + 0.013787181116640568, + 0.018506890162825584, + 0.015708431601524353, + 0.008165443316102028, + 0.005048162769526243, + 0.005482786800712347, + 0.0035515546333044767, + 0.011481836438179016, + 0.006712747737765312, + 0.006914319004863501, + 0.005741264205425978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.10871231555938721 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28134411573410034 + ], + "opposite_logprob_tokens": [ + -1.4063441753387451 + ], + "logprob_reward": -0.28134411573410034, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28134411573410034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18933053314685822, + 0.0887901559472084, + 0.08025495707988739, + 0.05470860376954079, + 0.04630211740732193, + 0.01585138961672783, + 0.04346524924039841, + 0.12360941618680954, + 0.32390618324279785, + 0.03378141671419144 + ], + "credit_map": [ + 0.027577532455325127, + 0.020154953002929688, + 0.020192638039588928, + 0.021480998024344444, + 0.032787974923849106, + 0.020097684115171432, + 0.026214497163891792, + 0.020824240520596504, + 0.013455779291689396, + 0.006591165903955698, + 0.010156777687370777, + 0.007141651585698128, + 0.008256847970187664, + 0.022506652399897575, + 0.011099977418780327, + 0.009581301361322403, + 0.009255321696400642, + 0.007036025170236826, + 0.008797508664429188, + 0.007507942616939545, + 0.013712105341255665, + 0.009993643499910831, + 0.01573723927140236, + 0.008215170353651047, + 0.004564643371850252, + 0.003993303515017033, + 0.005212094634771347, + 0.006396949291229248, + 0.006482590921223164, + 0.008334800601005554, + 0.012210788205265999, + 0.0075134350918233395, + 0.004023183137178421, + 0.0036342348903417587, + 0.0052320570684969425, + 0.007637340575456619, + 0.004545799922198057, + 0.004052184522151947, + 0.011134509928524494, + 0.006042803172022104, + 0.0014459948288276792, + 0.0018316798377782106, + 0.0018413623329252005, + 0.0021346842404454947, + 0.0027201552875339985, + 0.001908491482026875, + 0.0022402380127459764, + 0.0017287834780290723, + 0.002587377093732357, + 0.003559921169653535, + 0.008668212220072746, + 0.007298669777810574, + 0.00414765952154994, + 0.0056405384093523026, + 0.005414287559688091, + 0.006148584187030792, + 0.008499057963490486, + 0.012046447955071926, + 0.021280385553836823, + 0.012515654787421227, + 0.026371043175458908, + 0.013800512067973614, + 0.013663980178534985, + 0.015432329848408699, + 0.021017154678702354, + 0.053699884563684464, + 0.02701083943247795, + 0.032435204833745956, + 0.07959619164466858, + 0.026513783261179924, + 0.04242121800780296, + 0.04121189936995506, + 0.003042900236323476, + 0.0033186411019414663, + 0.0054297540336847305, + 0.003354760818183422, + 0.005113270599395037, + 0.0033983325120061636, + 0.005910919047892094, + 0.0042128367349505424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.28134411573410034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21467947959899902, + 0.087082639336586, + 0.03722758591175079, + 0.031612347811460495, + 0.11736834049224854, + 0.05368482694029808, + 0.032156381756067276, + 0.06776538491249084, + 0.24871942400932312, + 0.10970357805490494 + ], + "credit_map": [ + 0.03587333858013153, + 0.01587829925119877, + 0.01465498749166727, + 0.01615556702017784, + 0.026033269241452217, + 0.06198835000395775, + 0.030192352831363678, + 0.01390332542359829, + 0.012145682238042355, + 0.0049577513709664345, + 0.008708859793841839, + 0.005885336548089981, + 0.009876934811472893, + 0.02192167565226555, + 0.014266930520534515, + 0.009319469332695007, + 0.0045442963019013405, + 0.003132294164970517, + 0.005549546796828508, + 0.005679932422935963, + 0.004894273355603218, + 0.00485282251611352, + 0.004358586855232716, + 0.004215834196656942, + 0.0029252541717141867, + 0.002931920113041997, + 0.0031675405334681273, + 0.0037477680016309023, + 0.0048246379010379314, + 0.0028130982536822557, + 0.006986336782574654, + 0.004215792287141085, + 0.011426529847085476, + 0.007578828372061253, + 0.02374899946153164, + 0.01516406424343586, + 0.008812315762043, + 0.015537906438112259, + 0.02102297730743885, + 0.014076720923185349, + 0.0072303637862205505, + 0.005315745715051889, + 0.010219002142548561, + 0.005858156364411116, + 0.005033709574490786, + 0.0069048404693603516, + 0.006812118459492922, + 0.006310891825705767, + 0.002310435986146331, + 0.0025653494521975517, + 0.0033099520951509476, + 0.0067679425701498985, + 0.004632163792848587, + 0.0023553010541945696, + 0.0053731948137283325, + 0.0048420424573123455, + 0.0057420977391302586, + 0.007950435392558575, + 0.008510713465511799, + 0.007850443944334984, + 0.010177246294915676, + 0.008862034417688847, + 0.010768893174827099, + 0.007903522811830044, + 0.015581408515572548, + 0.02305319905281067, + 0.022879116237163544, + 0.043066948652267456, + 0.027788974344730377, + 0.024436507374048233, + 0.06106976792216301, + 0.03084350936114788, + 0.014873379841446877, + 0.01449095830321312, + 0.012630272656679153, + 0.013392288237810135, + 0.013644387014210224, + 0.019636178389191628, + 0.01034192368388176, + 0.010694198310375214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11252021044492722, + 0.09792984277009964, + 0.26429152488708496, + 0.07223628461360931, + 0.1744876652956009, + 0.0778469517827034, + 0.04926856607198715, + 0.060381535440683365, + 0.07273942977190018, + 0.018297940492630005 + ], + "credit_map": [ + 0.01306564174592495, + 0.009683267213404179, + 0.021829789504408836, + 0.01542973704636097, + 0.012256945483386517, + 0.01475598942488432, + 0.012712275609374046, + 0.012786569073796272, + 0.01062376331537962, + 0.008519955910742283, + 0.01891847513616085, + 0.012371605262160301, + 0.011115836910903454, + 0.014269895851612091, + 0.010786228813230991, + 0.011324092745780945, + 0.04314878210425377, + 0.03984835371375084, + 0.03438546136021614, + 0.030620897188782692, + 0.03528653457760811, + 0.027898656204342842, + 0.028198959305882454, + 0.024903878569602966, + 0.007528114598244429, + 0.0060695079155266285, + 0.009844237007200718, + 0.006899785716086626, + 0.007179673295468092, + 0.015073776245117188, + 0.009646705351769924, + 0.009994489140808582, + 0.011580417864024639, + 0.006897874642163515, + 0.06329086422920227, + 0.015680693089962006, + 0.008960641920566559, + 0.025880001485347748, + 0.0158538855612278, + 0.02634331025183201, + 0.008293972350656986, + 0.006572318729013205, + 0.017847491428256035, + 0.007038367446511984, + 0.009777780622243881, + 0.01154501922428608, + 0.006854277104139328, + 0.009917733259499073, + 0.007027550134807825, + 0.004345664754509926, + 0.008721769787371159, + 0.005269536282867193, + 0.0051943473517894745, + 0.007829776033759117, + 0.0050101811066269875, + 0.005869742948561907, + 0.005180476699024439, + 0.004215385764837265, + 0.009979303926229477, + 0.008369307965040207, + 0.008394704200327396, + 0.008564217016100883, + 0.007714080158621073, + 0.007964065298438072, + 0.008555015549063683, + 0.006537078879773617, + 0.010605056770145893, + 0.009887873195111752, + 0.008284162729978561, + 0.008356449194252491, + 0.011557947844266891, + 0.008955846540629864, + 0.002954753814265132, + 0.0018204208463430405, + 0.0019903837237507105, + 0.0016872006235644221, + 0.003034410532563925, + 0.0028818617574870586, + 0.0019257907988503575, + 0.0020031193271279335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character wearing a period costume riding a chestnut horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character's long dark brown hair fluttering violently in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a faded blue-green coarse coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an endless expanse of golden sand dunes with a massive shifting sand castle in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Ghibli animation style with saturated soft colors, a hand-painted washed blue sky, and fluffy white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319446949288249 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.00319446949288249, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319446949288249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31853270530700684, + 0.03820967674255371, + 0.17903077602386475, + 0.06870771944522858, + 0.07402278482913971, + 0.09208553284406662, + 0.03691190108656883, + 0.0729576125741005, + 0.07472552359104156, + 0.04481586068868637 + ], + "credit_map": [ + 0.05637718737125397, + 0.034449897706508636, + 0.04114776477217674, + 0.034884583204984665, + 0.03069416992366314, + 0.058424822986125946, + 0.02794903703033924, + 0.03460520878434181, + 0.004321751184761524, + 0.0032963494304567575, + 0.007013513240963221, + 0.004505746066570282, + 0.0038210581988096237, + 0.005588977597653866, + 0.004960149992257357, + 0.004702126141637564, + 0.030581455677747726, + 0.026220131665468216, + 0.02775472030043602, + 0.023290492594242096, + 0.022021668031811714, + 0.01823381334543228, + 0.013285214081406593, + 0.017643269151449203, + 0.004318677820265293, + 0.004604136571288109, + 0.005103897303342819, + 0.009664122946560383, + 0.00822401512414217, + 0.008810432627797127, + 0.017048871144652367, + 0.010933552868664265, + 0.008906176313757896, + 0.00537378853186965, + 0.009784823283553123, + 0.010528868064284325, + 0.005897426977753639, + 0.01096176728606224, + 0.013864267617464066, + 0.00870566163212061, + 0.01260768435895443, + 0.009868570603430271, + 0.01887851022183895, + 0.009531284682452679, + 0.007238057907670736, + 0.012496194802224636, + 0.010352381505072117, + 0.011112849228084087, + 0.0026722783222794533, + 0.003270649118348956, + 0.006360105238854885, + 0.005166364833712578, + 0.005083474330604076, + 0.004437150899320841, + 0.005432503763586283, + 0.004489372484385967, + 0.00775794917717576, + 0.0060274843126535416, + 0.012381733395159245, + 0.009145739488303661, + 0.008272575214505196, + 0.01272718608379364, + 0.008062024600803852, + 0.00858291331678629, + 0.005000497214496136, + 0.011339171789586544, + 0.007850578054785728, + 0.009309518150985241, + 0.013130380772054195, + 0.006245824508368969, + 0.01264365203678608, + 0.00920589454472065, + 0.0071886638179421425, + 0.003960380796343088, + 0.00676505733281374, + 0.0032303736079484224, + 0.007336942013353109, + 0.007061128970235586, + 0.004180310759693384, + 0.005092998500913382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.00319446949288249 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_tensor([5021, 5021])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_tensor([5021, 5021])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7fe7ab5dab2445eb13db0b744c3c51cdb42a5a72 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_tensor([5021, 5021])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic long shot sweeps across a vast, sun-drenched meadow in a slow, graceful arc. At the center of this idyllic scene, a woman with long, flowing auburn hair twirls effortlessly. She wears a vibrant, long, earth-yellow dress crafted from delicate, lightweight chiffon that catches the breeze. As she steadily spins on the soles of her feet, her arms extended for balance, her head tilted back in blissful abandon, a serene smile on her face. The vibrant yellow fabric of her dress forms a perfect circle around her, creating a striking visual contrast with the lush green grass dotted with small wildflowers. The entire scene is bathed in the warm golden glow of the afternoon sun, which, as the camera pans, casts long, soft shadows and creates cinematic lens flares, backlighting her body and making the translucent material of her dress shimmer. The atmosphere is one of pure, unadulterated joy and liberated tranquility, capturing a fleeting, dreamlike moment of perfect freedom.", + "seed": 20809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_5021_seed_20809_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009172456344822421, + "vlm_logprob": -0.009172456344822421, + "vlm_yes_no_margin": 5.775, + "vlm_token_credit": -0.009172456344822421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006825820542871952 + ], + "opposite_logprob_tokens": [ + -5.631825923919678 + ], + "logprob_reward": -0.006825820542871952, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006825820542871952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10789307951927185, + 0.020420977845788002, + 0.023276112973690033, + 0.3568304479122162, + 0.08514026552438736, + 0.024308469146490097, + 0.05924718827009201, + 0.04619939625263214, + 0.08127202093601227, + 0.19541214406490326 + ], + "credit_map": [ + 0.009487233124673367, + 0.01147286593914032, + 0.016957446932792664, + 0.017417434602975845, + 0.012315687723457813, + 0.0140694510191679, + 0.011869268491864204, + 0.014303688891232014, + 0.0018538790754973888, + 0.0022232038900256157, + 0.0021099159494042397, + 0.002912951400503516, + 0.0032030625734478235, + 0.002020832384005189, + 0.0024349037557840347, + 0.0036622262559831142, + 0.002687386702746153, + 0.0018091804813593626, + 0.002716659801080823, + 0.00356678431853652, + 0.002152644330635667, + 0.003458147868514061, + 0.003339360235258937, + 0.0035459476057440042, + 0.023684533312916756, + 0.008546443656086922, + 0.019920261576771736, + 0.03272883594036102, + 0.012888018041849136, + 0.08496473729610443, + 0.03302568942308426, + 0.1410718858242035, + 0.012207618914544582, + 0.004842430353164673, + 0.03243667632341385, + 0.005596430506557226, + 0.004710542503744364, + 0.008632747456431389, + 0.008281233720481396, + 0.008432581089437008, + 0.004768806044012308, + 0.002125733532011509, + 0.005323454737663269, + 0.0017019854858517647, + 0.0022086109966039658, + 0.002617952646687627, + 0.0022647245787084103, + 0.0032971990294754505, + 0.012644748203456402, + 0.006924123037606478, + 0.008646183647215366, + 0.005524461157619953, + 0.007748091593384743, + 0.0046678576618433, + 0.007121817208826542, + 0.005969899240881205, + 0.0034534018486738205, + 0.0030060280114412308, + 0.012164944782853127, + 0.006362728774547577, + 0.005390845239162445, + 0.003703504567965865, + 0.006189373787492514, + 0.005928566213697195, + 0.008979431353509426, + 0.005236473400145769, + 0.028425799682736397, + 0.007869144901633263, + 0.003895112778991461, + 0.00792703777551651, + 0.007183547597378492, + 0.011755461804568768, + 0.01399634312838316, + 0.013128994032740593, + 0.03981930762529373, + 0.028506027534604073, + 0.014483818784356117, + 0.023073282092809677, + 0.028261356055736542, + 0.034143004566431046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.006825820542871952 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013487775577232242 + ], + "opposite_logprob_tokens": [ + -6.626348972320557 + ], + "logprob_reward": -0.0013487775577232242, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013487775577232242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0786978229880333, + 0.04586736112833023, + 0.05037957802414894, + 0.33041298389434814, + 0.2935355007648468, + 0.015343447215855122, + 0.024407146498560905, + 0.036015983670949936, + 0.08270925283432007, + 0.04263092577457428 + ], + "credit_map": [ + 0.008151954971253872, + 0.008032926358282566, + 0.009886018000543118, + 0.010265018790960312, + 0.009726610034704208, + 0.013021735474467278, + 0.009490929543972015, + 0.0101226307451725, + 0.0032720104791224003, + 0.004423385951668024, + 0.0060579413548111916, + 0.005841840524226427, + 0.009946822188794613, + 0.00388899608515203, + 0.005346309393644333, + 0.007090055383741856, + 0.004428398795425892, + 0.004067213274538517, + 0.0068472339771687984, + 0.006500055082142353, + 0.00740901380777359, + 0.005009017884731293, + 0.0073648677207529545, + 0.00875377468764782, + 0.022530272603034973, + 0.00668867863714695, + 0.01713215373456478, + 0.02962036244571209, + 0.008970262482762337, + 0.07731018215417862, + 0.028170069679617882, + 0.13999100029468536, + 0.04036036133766174, + 0.01180325634777546, + 0.12128842622041702, + 0.01991794817149639, + 0.014698115177452564, + 0.031056610867381096, + 0.028053244575858116, + 0.026357537135481834, + 0.001780167338438332, + 0.0011940845288336277, + 0.0038898629136383533, + 0.001016671652905643, + 0.0018163309432566166, + 0.0019170112209394574, + 0.001448439434170723, + 0.0022808799985796213, + 0.004228661302477121, + 0.002758124377578497, + 0.003999576438218355, + 0.002444879850372672, + 0.0034932876005768776, + 0.0021605966612696648, + 0.0027211299166083336, + 0.0026008915156126022, + 0.0021368814632296562, + 0.002153946552425623, + 0.008852049708366394, + 0.00256911339238286, + 0.004411512054502964, + 0.0030255375895649195, + 0.0032160887494683266, + 0.009650852531194687, + 0.007431312929838896, + 0.01379381399601698, + 0.012113797478377819, + 0.00981298927217722, + 0.005881522316485643, + 0.008174543268978596, + 0.01317532453685999, + 0.012325949966907501, + 0.0031092208810150623, + 0.004113849718123674, + 0.007743378169834614, + 0.006935155484825373, + 0.004460903350263834, + 0.004089502617716789, + 0.005849661771208048, + 0.0063292523846030235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.0013487775577232242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03209064155817032, + 0.023616496473550797, + 0.09799911081790924, + 0.21526102721691132, + 0.4985870122909546, + 0.012128669768571854, + 0.013736252672970295, + 0.02356667071580887, + 0.04235980659723282, + 0.04065437242388725 + ], + "credit_map": [ + 0.003372121136635542, + 0.0036878071259707212, + 0.0045312498696148396, + 0.004026775248348713, + 0.003396035172045231, + 0.005620415788143873, + 0.00372806703671813, + 0.003728169947862625, + 0.0018322571413591504, + 0.001665004063397646, + 0.002342653926461935, + 0.003871018998324871, + 0.003068168880417943, + 0.0018755183555185795, + 0.003864513011649251, + 0.005097359884530306, + 0.007980115711688995, + 0.004939850885421038, + 0.008397093042731285, + 0.013373102061450481, + 0.006839729845523834, + 0.019573312252759933, + 0.01029796339571476, + 0.02659793756902218, + 0.014363429509103298, + 0.0045001693069934845, + 0.010599648579955101, + 0.01947147585451603, + 0.006079066079109907, + 0.05075439438223839, + 0.01827055774629116, + 0.09122227132320404, + 0.07043563574552536, + 0.021015863865613937, + 0.206198588013649, + 0.03370578587055206, + 0.024336589500308037, + 0.05157402157783508, + 0.046787794679403305, + 0.0445326492190361, + 0.001381936832331121, + 0.0010259903501719236, + 0.0026102673728019, + 0.0012317050714045763, + 0.0011641725432127714, + 0.0016038145404309034, + 0.001262694364413619, + 0.0018480878788977861, + 0.0014781104400753975, + 0.0011769927805289626, + 0.0025842830073088408, + 0.0016261524287983775, + 0.0025120258796960115, + 0.0012882313458248973, + 0.0017183009767904878, + 0.0013521539513021708, + 0.001464253175072372, + 0.0014837814960628748, + 0.005622313357889652, + 0.0020831774454563856, + 0.0036283761728554964, + 0.0017019821098074317, + 0.0022033697459846735, + 0.005379415117204189, + 0.0038019162602722645, + 0.004659425467252731, + 0.010841386392712593, + 0.003060366725549102, + 0.002173652872443199, + 0.008678439073264599, + 0.0034241415560245514, + 0.005720480810850859, + 0.0022311792708933353, + 0.0038602245040237904, + 0.0053683859296143055, + 0.00843842327594757, + 0.0037782916333526373, + 0.0033689250703901052, + 0.008132378570735455, + 0.005476559977978468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786981642246246 + ], + "opposite_logprob_tokens": [ + -3.529787063598633 + ], + "logprob_reward": -0.029786981642246246, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786981642246246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04184512421488762, + 0.022223344072699547, + 0.021884920075535774, + 0.42056408524513245, + 0.33395612239837646, + 0.018344517797231674, + 0.019305603578686714, + 0.039209190756082535, + 0.05084709823131561, + 0.031819988042116165 + ], + "credit_map": [ + 0.0031337060499936342, + 0.004914498887956142, + 0.0063936770893633366, + 0.009731991216540337, + 0.0036627340596169233, + 0.003666915465146303, + 0.005102249793708324, + 0.005239352583885193, + 0.001654179417528212, + 0.0016687361057847738, + 0.0021263454109430313, + 0.0033637722954154015, + 0.004546557553112507, + 0.0016391131794080138, + 0.0030349299777299166, + 0.004189711529761553, + 0.0024833420757204294, + 0.0014641456073150039, + 0.0029479030054062605, + 0.003011514199897647, + 0.0021079040598124266, + 0.003077029949054122, + 0.0026485342532396317, + 0.0041445461101830006, + 0.028118861839175224, + 0.008086352609097958, + 0.020116662606596947, + 0.038062918931245804, + 0.011084901168942451, + 0.09792756289243698, + 0.03631201013922691, + 0.18085479736328125, + 0.04510170593857765, + 0.01339541096240282, + 0.1381368190050125, + 0.023007795214653015, + 0.0189357902854681, + 0.033762987703084946, + 0.03242233395576477, + 0.029193280264735222, + 0.00215318170376122, + 0.0017557214014232159, + 0.004680299200117588, + 0.0015845086891204119, + 0.0019714399240911007, + 0.002234680112451315, + 0.002000344917178154, + 0.001964343711733818, + 0.003397219581529498, + 0.0016251736087724566, + 0.003333204658702016, + 0.001842178520746529, + 0.002751026302576065, + 0.0019414029084146023, + 0.002010985976085067, + 0.0024044120218604803, + 0.0031920296605676413, + 0.0028929063118994236, + 0.007095710840076208, + 0.002708521205931902, + 0.003276721341535449, + 0.0039937542751431465, + 0.003554644761607051, + 0.012494903989136219, + 0.005122474394738674, + 0.004468002822250128, + 0.01260213553905487, + 0.005300363525748253, + 0.0025971075519919395, + 0.0056447479873895645, + 0.0077893552370369434, + 0.007322908379137516, + 0.0016699857078492641, + 0.002916663885116577, + 0.005435872357338667, + 0.006284281611442566, + 0.0028065068181604147, + 0.0028585016261786222, + 0.0059164101257920265, + 0.0039317673072218895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.029786981642246246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002706876548472792 + ], + "opposite_logprob_tokens": [ + -8.25027084350586 + ], + "logprob_reward": -0.0002706876548472792, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0002706876548472792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07245111465454102, + 0.03213878720998764, + 0.024494031444191933, + 0.15119215846061707, + 0.09687122702598572, + 0.04074927419424057, + 0.05342442914843559, + 0.20963549613952637, + 0.17727956175804138, + 0.1417638659477234 + ], + "credit_map": [ + 0.0064467559568583965, + 0.008421467617154121, + 0.010666263289749622, + 0.010968430899083614, + 0.008644427172839642, + 0.006360925268381834, + 0.011464101262390614, + 0.009478743188083172, + 0.002733828267082572, + 0.003159146523103118, + 0.004021613858640194, + 0.00637689046561718, + 0.00569766154512763, + 0.0019139586947858334, + 0.004186516627669334, + 0.004049170296639204, + 0.0024568403605371714, + 0.0015112411929294467, + 0.0029086510185152292, + 0.0034863571636378765, + 0.0021292930468916893, + 0.0037465100176632404, + 0.0034308661706745625, + 0.004824275150895119, + 0.010440325364470482, + 0.0033615378197282553, + 0.008333833888173103, + 0.013779133558273315, + 0.004626249894499779, + 0.03378596529364586, + 0.013512767851352692, + 0.06335235387086868, + 0.010714824311435223, + 0.004046745132654905, + 0.03728447109460831, + 0.00807923823595047, + 0.008099699392914772, + 0.010744384489953518, + 0.00995261687785387, + 0.00794924609363079, + 0.006005601957440376, + 0.003414016216993332, + 0.006404608488082886, + 0.004682409577071667, + 0.004517502151429653, + 0.003381640650331974, + 0.005509349517524242, + 0.0068341465666890144, + 0.004781818483024836, + 0.003768142545595765, + 0.011163868010044098, + 0.005663819145411253, + 0.010136590339243412, + 0.004725403152406216, + 0.006666539702564478, + 0.006518248002976179, + 0.014521828852593899, + 0.011968536302447319, + 0.06732993572950363, + 0.011467128992080688, + 0.009185932576656342, + 0.017609059810638428, + 0.013379266485571861, + 0.06417380273342133, + 0.022040409967303276, + 0.017946554347872734, + 0.025513136759400368, + 0.01746649108827114, + 0.00918702594935894, + 0.01722019352018833, + 0.022339193150401115, + 0.04556656256318092, + 0.009104238823056221, + 0.01337724830955267, + 0.030685879290103912, + 0.02106369659304619, + 0.012321937829256058, + 0.013568050228059292, + 0.023461923003196716, + 0.018180888146162033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.0002706876548472792 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic long shot sweeps across a vast, sun-drenched meadow in a slow, graceful arc. At the center of this idyllic scene, a woman with long, flowing auburn hair twirls effortlessly. She wears a vibrant, long, earth-yellow dress crafted from delicate, lightweight chiffon that catches the breeze. As she steadily spins on the soles of her feet, her arms extended for balance, her head tilted back in blissful abandon, a serene smile on her face. The vibrant yellow fabric of her dress forms a perfect circle around her, creating a striking visual contrast with the lush green grass dotted with small wildflowers. The entire scene is bathed in the warm golden glow of the afternoon sun, which, as the camera pans, casts long, soft shadows and creates cinematic lens flares, backlighting her body and making the translucent material of her dress shimmer. The atmosphere is one of pure, unadulterated joy and liberated tranquility, capturing a fleeting, dreamlike moment of perfect freedom.", + "seed": 20810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_5021_seed_20810_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014324565016431734, + "vlm_logprob": -0.014324565016431734, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.014324565016431734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009569723159074783 + ], + "opposite_logprob_tokens": [ + -5.0095696449279785 + ], + "logprob_reward": -0.009569723159074783, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009569723159074783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5359355807304382, + 0.08386243134737015, + 0.05172999948263168, + 0.14121663570404053, + 0.0343620739877224, + 0.033013250678777695, + 0.05863269418478012, + 0.03149566799402237, + 0.01423728559166193, + 0.015514387749135494 + ], + "credit_map": [ + 0.06894948333501816, + 0.0216219425201416, + 0.16859988868236542, + 0.051564425230026245, + 0.02582816779613495, + 0.032598916441202164, + 0.04909919202327728, + 0.11767355352640152, + 0.010390113107860088, + 0.005410612560808659, + 0.023490073159337044, + 0.006119330879300833, + 0.005801532883197069, + 0.005663941614329815, + 0.005854051094502211, + 0.021132763475179672, + 0.005453688092529774, + 0.004854596685618162, + 0.009063798002898693, + 0.004682908300310373, + 0.006209411658346653, + 0.007082547526806593, + 0.004102072678506374, + 0.010280974209308624, + 0.01375318318605423, + 0.006011392921209335, + 0.04924694448709488, + 0.011186642572283745, + 0.0076176561415195465, + 0.006834967527538538, + 0.010702824220061302, + 0.03586301580071449, + 0.002585117006674409, + 0.003023793688043952, + 0.006172744557261467, + 0.005349228624254465, + 0.004146350082010031, + 0.002954864175990224, + 0.005228155292570591, + 0.004901818465441465, + 0.004042677581310272, + 0.0014110059710219502, + 0.01151882205158472, + 0.0021615743171423674, + 0.0019217951921746135, + 0.002897194353863597, + 0.0015682056546211243, + 0.007491974160075188, + 0.009175417944788933, + 0.006395831238478422, + 0.008716639131307602, + 0.004654861520975828, + 0.004133548121899366, + 0.006662546191364527, + 0.004033771343529224, + 0.014860072173178196, + 0.003765442641451955, + 0.004127883352339268, + 0.004822847433388233, + 0.0031103831715881824, + 0.0026636391412466764, + 0.003688908414915204, + 0.003514905460178852, + 0.005801656749099493, + 0.0015095879789441824, + 0.0012769994791597128, + 0.0017811994766816497, + 0.0024409573525190353, + 0.0021492992527782917, + 0.0010490501299500465, + 0.002443131059408188, + 0.001587058650329709, + 0.0016752578085288405, + 0.0016504355007782578, + 0.0019052185816690326, + 0.002151849213987589, + 0.0018528263317421079, + 0.0012068393407389522, + 0.0024657200556248426, + 0.0026062396354973316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009569723159074783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007295330869965255 + ], + "opposite_logprob_tokens": [ + -7.250729560852051 + ], + "logprob_reward": -0.0007295330869965255, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007295330869965255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.767238199710846, + 0.031019430607557297, + 0.023761361837387085, + 0.03294142708182335, + 0.08344124257564545, + 0.007146546617150307, + 0.008296866901218891, + 0.00981942843645811, + 0.011616767384111881, + 0.02471863478422165 + ], + "credit_map": [ + 0.09508651494979858, + 0.02877231314778328, + 0.24268615245819092, + 0.07088316977024078, + 0.0341833159327507, + 0.04579653963446617, + 0.07662034034729004, + 0.17320992052555084, + 0.004553527105599642, + 0.0023539168760180473, + 0.008779131807386875, + 0.0030518618877977133, + 0.0035338702145963907, + 0.0022467770613729954, + 0.003441754961386323, + 0.003058593487367034, + 0.0023006731644272804, + 0.002684576204046607, + 0.004727771505713463, + 0.0029951923061162233, + 0.002906502690166235, + 0.002245117211714387, + 0.0024123480543494225, + 0.003489184658974409, + 0.0032492156606167555, + 0.001476885168813169, + 0.011679473333060741, + 0.00161495350766927, + 0.0021794354543089867, + 0.0018570899264886975, + 0.0017380304634571075, + 0.009146342985332012, + 0.0068862526677548885, + 0.005001784302294254, + 0.016800418496131897, + 0.017769383266568184, + 0.004901532083749771, + 0.00568264489993453, + 0.009393750689923763, + 0.01700548268854618, + 0.0009215260506607592, + 0.0008250065729953349, + 0.001480503473430872, + 0.0012262017698958516, + 0.00085446861339733, + 0.0005327058606781065, + 0.0005882316618226469, + 0.000717903720214963, + 0.001304354751482606, + 0.0010603897972032428, + 0.000992812099866569, + 0.000828527205158025, + 0.001068238401785493, + 0.000996356480754912, + 0.0010114414617419243, + 0.0010347473435103893, + 0.0012817146489396691, + 0.0014823195524513721, + 0.0010780792217701674, + 0.0011614931281656027, + 0.0012407316826283932, + 0.0009853772353380919, + 0.0014876177301630378, + 0.0011020962847396731, + 0.0011250011157244444, + 0.0011914734495803714, + 0.0018861573189496994, + 0.001593732158653438, + 0.0016478552715852857, + 0.0007235546363517642, + 0.0018176852026954293, + 0.0016313096275553107, + 0.0022473435383290052, + 0.0023080729879438877, + 0.003575453767552972, + 0.003678301814943552, + 0.004571197088807821, + 0.0016291532665491104, + 0.003878024872392416, + 0.0028310890775173903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0007295330869965255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950473714619875 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950473714619875, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950473714619875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5510914325714111, + 0.10854402929544449, + 0.0655069425702095, + 0.06386838108301163, + 0.14297586679458618, + 0.015798082575201988, + 0.007844340056180954, + 0.012009796686470509, + 0.02096114307641983, + 0.011400007642805576 + ], + "credit_map": [ + 0.0683836117386818, + 0.02122361958026886, + 0.17259155213832855, + 0.051639117300510406, + 0.025705354288220406, + 0.03318120166659355, + 0.05785323679447174, + 0.12051371484994888, + 0.014074952341616154, + 0.004352817311882973, + 0.040988050401210785, + 0.0038689388893544674, + 0.0036383792757987976, + 0.005905186291784048, + 0.004986971616744995, + 0.030728735029697418, + 0.007218056358397007, + 0.006053696386516094, + 0.008753886446356773, + 0.01177445612847805, + 0.008783290162682533, + 0.006126685533672571, + 0.008449195884168148, + 0.008347675204277039, + 0.00783058162778616, + 0.006413010880351067, + 0.010797442868351936, + 0.008781644515693188, + 0.007509293500334024, + 0.005451819859445095, + 0.010476046241819859, + 0.0066085453145205975, + 0.010846748948097229, + 0.008583899587392807, + 0.030783796682953835, + 0.029834263026714325, + 0.008148347027599812, + 0.009436782449483871, + 0.016441194340586662, + 0.028900831937789917, + 0.0017812995938584208, + 0.0007616716902703047, + 0.005456126295030117, + 0.0009332791087217629, + 0.0009700015070848167, + 0.001196320285089314, + 0.0007640193798579276, + 0.003935364540666342, + 0.0013262004358693957, + 0.0010065429378300905, + 0.0009583585779182613, + 0.0009232290904037654, + 0.0010321969166398048, + 0.0006739809759892523, + 0.000970145280007273, + 0.0009536863653920591, + 0.00101738260127604, + 0.0012480667792260647, + 0.0014971316559240222, + 0.0018737106584012508, + 0.0016699840780347586, + 0.0011087435996159911, + 0.0021170068066567183, + 0.0014777694595977664, + 0.0017625503242015839, + 0.0018767223227769136, + 0.003380456706508994, + 0.0034615593031048775, + 0.003601532429456711, + 0.0014374088495969772, + 0.0034059120807796717, + 0.0020350010599941015, + 0.0009508017683401704, + 0.0011955185327678919, + 0.0014193368842825294, + 0.0019394435221329331, + 0.0016332019586116076, + 0.0008827279671095312, + 0.0015059186844155192, + 0.0018730578012764454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005950473714619875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493330955505371 + ], + "opposite_logprob_tokens": [ + -2.9299333095550537 + ], + "logprob_reward": -0.05493330955505371, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493330955505371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8320783972740173, + 0.018273450434207916, + 0.021122492849826813, + 0.04180697724223137, + 0.013374602422118187, + 0.015871455892920494, + 0.016620244830846786, + 0.005459302105009556, + 0.026249118149280548, + 0.009143900126218796 + ], + "credit_map": [ + 0.1033235639333725, + 0.03059607371687889, + 0.2662976086139679, + 0.07634756714105606, + 0.03531752526760101, + 0.049780167639255524, + 0.08038751035928726, + 0.19002842903137207, + 0.002594331046566367, + 0.0019105625106021762, + 0.002719484968110919, + 0.0022711188066750765, + 0.0028210452292114496, + 0.0014855953631922603, + 0.002498934045433998, + 0.0019723798613995314, + 0.0024751760065555573, + 0.002102895174175501, + 0.0037274518981575966, + 0.0030169147066771984, + 0.002855523955076933, + 0.002001771703362465, + 0.0022133691236376762, + 0.0027293935418128967, + 0.004960878286510706, + 0.004279096145182848, + 0.005141942296177149, + 0.005955272354185581, + 0.005905811674892902, + 0.003883574390783906, + 0.00801183469593525, + 0.0036685720551759005, + 0.0015230418648570776, + 0.000958697113674134, + 0.0022023983765393496, + 0.002376304240897298, + 0.0011671006213873625, + 0.0010196483926847577, + 0.0023597560357302427, + 0.0017676568822935224, + 0.0018764663254842162, + 0.001542978803627193, + 0.003613189561292529, + 0.0018153548007830977, + 0.00125926174223423, + 0.0015648907283321023, + 0.0009696231572888792, + 0.003229692345485091, + 0.0022690976038575172, + 0.0016312244115397334, + 0.003773374715819955, + 0.0012158445315435529, + 0.0013884158106520772, + 0.0011555299861356616, + 0.0012737954966723919, + 0.003912962973117828, + 0.0006777310627512634, + 0.0007593246991746128, + 0.0008632427779957652, + 0.0005298287724144757, + 0.00042948793270625174, + 0.0005934307700954378, + 0.0006467101629823446, + 0.0009595465962775052, + 0.001890963059850037, + 0.0023570663761347532, + 0.004664174281060696, + 0.0032421997748315334, + 0.00426879758015275, + 0.0017184694297611713, + 0.004441611934453249, + 0.0036658381577581167, + 0.0007800667081028223, + 0.0008397154160775244, + 0.0013662048149853945, + 0.001364607596769929, + 0.0015263272216543555, + 0.0005927963647991419, + 0.001484901993535459, + 0.0011892813490703702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.05493330955505371 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004397855664137751 + ], + "opposite_logprob_tokens": [ + -7.750439643859863 + ], + "logprob_reward": -0.0004397855664137751, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004397855664137751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5833517909049988, + 0.08683650940656662, + 0.03915151581168175, + 0.08704965561628342, + 0.061380643397569656, + 0.05631589889526367, + 0.02193318121135235, + 0.012406956404447556, + 0.02400895208120346, + 0.027564942836761475 + ], + "credit_map": [ + 0.07231836766004562, + 0.02191009372472763, + 0.18427544832229614, + 0.05556482449173927, + 0.025840040296316147, + 0.034574564546346664, + 0.05941937863826752, + 0.1294490545988083, + 0.00993955135345459, + 0.010246235877275467, + 0.00976574793457985, + 0.01458919420838356, + 0.012551242485642433, + 0.006647823844105005, + 0.012319658882915974, + 0.01077705156058073, + 0.0028462507762014866, + 0.004426708910614252, + 0.007466600742191076, + 0.005636319518089294, + 0.0052804104052484035, + 0.0027150448877364397, + 0.003649406600743532, + 0.007130772806704044, + 0.011349983513355255, + 0.00560341589152813, + 0.02478651888668537, + 0.00987514853477478, + 0.006552122067660093, + 0.0054998453706502914, + 0.009431292302906513, + 0.013951323926448822, + 0.005392394494265318, + 0.004528867546468973, + 0.012444548308849335, + 0.012416614219546318, + 0.004416940733790398, + 0.00453155254945159, + 0.007580389268696308, + 0.010069334879517555, + 0.010991753078997135, + 0.008556152693927288, + 0.009577993303537369, + 0.008005077950656414, + 0.005179553292691708, + 0.005489788483828306, + 0.0037663306575268507, + 0.004749245010316372, + 0.00279574329033494, + 0.0022481728810817003, + 0.0031237981747835875, + 0.0017279337625950575, + 0.0025114223826676607, + 0.001932354993186891, + 0.002755016554147005, + 0.004838735796511173, + 0.001832261448726058, + 0.0018091173842549324, + 0.001564273377880454, + 0.000983673962764442, + 0.0010712523944675922, + 0.001491404022090137, + 0.0012445416068658233, + 0.0024104320909827948, + 0.0016583980759605765, + 0.0019077655160799623, + 0.003346468089148402, + 0.004208825528621674, + 0.004463461693376303, + 0.001531361136585474, + 0.004276348743587732, + 0.002616320736706257, + 0.002569901756942272, + 0.002882658736780286, + 0.004438150208443403, + 0.0043416996486485004, + 0.003224446903914213, + 0.0024002306163311005, + 0.00367186963558197, + 0.0040359823033213615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0004397855664137751 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic long shot sweeps across a vast, sun-drenched meadow in a slow, graceful arc. At the center of this idyllic scene, a woman with long, flowing auburn hair twirls effortlessly. She wears a vibrant, long, earth-yellow dress crafted from delicate, lightweight chiffon that catches the breeze. As she steadily spins on the soles of her feet, her arms extended for balance, her head tilted back in blissful abandon, a serene smile on her face. The vibrant yellow fabric of her dress forms a perfect circle around her, creating a striking visual contrast with the lush green grass dotted with small wildflowers. The entire scene is bathed in the warm golden glow of the afternoon sun, which, as the camera pans, casts long, soft shadows and creates cinematic lens flares, backlighting her body and making the translucent material of her dress shimmer. The atmosphere is one of pure, unadulterated joy and liberated tranquility, capturing a fleeting, dreamlike moment of perfect freedom.", + "seed": 20811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_5021_seed_20811_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03014886719174683, + "vlm_logprob": -0.03014886719174683, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.03014886719174683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02613915130496025 + ], + "opposite_logprob_tokens": [ + -3.776139259338379 + ], + "logprob_reward": -0.02613915130496025, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02613915130496025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19256259500980377, + 0.10317967087030411, + 0.0610516183078289, + 0.07737226039171219, + 0.0651971772313118, + 0.10423888266086578, + 0.05739691108465195, + 0.22415941953659058, + 0.050957415252923965, + 0.06388404220342636 + ], + "credit_map": [ + 0.030011916533112526, + 0.029757443815469742, + 0.041769418865442276, + 0.013389811851084232, + 0.010514521971344948, + 0.03545239940285683, + 0.011896156705915928, + 0.019770922139286995, + 0.020782658830285072, + 0.020231161266565323, + 0.019808201119303703, + 0.004635739140212536, + 0.006009295582771301, + 0.019466714933514595, + 0.004534637089818716, + 0.007711266167461872, + 0.007848840206861496, + 0.013836294412612915, + 0.009658582508563995, + 0.004688831977546215, + 0.003844328923150897, + 0.01186391618102789, + 0.005296417977660894, + 0.004014411009848118, + 0.01054272148758173, + 0.012789028696715832, + 0.018687911331653595, + 0.00773271219804883, + 0.004105119500309229, + 0.009736383333802223, + 0.007903045043349266, + 0.005875337403267622, + 0.013911750167608261, + 0.00871750432997942, + 0.01109058503061533, + 0.005282845348119736, + 0.005409437231719494, + 0.007450613658875227, + 0.007318142801523209, + 0.006016300525516272, + 0.012296615168452263, + 0.02177198976278305, + 0.025396285578608513, + 0.005592741072177887, + 0.007827776484191418, + 0.018329421058297157, + 0.008170902729034424, + 0.0048531536012887955, + 0.00952904298901558, + 0.0130448117852211, + 0.00829948578029871, + 0.003559903474524617, + 0.003929986618459225, + 0.009103180840611458, + 0.00580243393778801, + 0.0041280630975961685, + 0.04611988365650177, + 0.028602533042430878, + 0.036679986864328384, + 0.014544583857059479, + 0.013647236861288548, + 0.025479618459939957, + 0.01716659590601921, + 0.04191899299621582, + 0.007817539386451244, + 0.008773209527134895, + 0.008162095211446285, + 0.0042594559490680695, + 0.0036698132753372192, + 0.007245596963912249, + 0.0049445126205682755, + 0.0060851965099573135, + 0.009907584637403488, + 0.008588732220232487, + 0.006443213205784559, + 0.00555030070245266, + 0.01095225103199482, + 0.008497712202370167, + 0.0073246704414486885, + 0.006619584280997515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.02613915130496025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001347706071101129 + ], + "opposite_logprob_tokens": [ + -6.626347541809082 + ], + "logprob_reward": -0.001347706071101129, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001347706071101129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1689731329679489, + 0.07979185879230499, + 0.11017069965600967, + 0.05554591119289398, + 0.0450969934463501, + 0.12386199086904526, + 0.06857196241617203, + 0.183211550116539, + 0.04250587150454521, + 0.12227007746696472 + ], + "credit_map": [ + 0.022211959585547447, + 0.02291947603225708, + 0.028852440416812897, + 0.014570794999599457, + 0.00949798058718443, + 0.03656104952096939, + 0.01845739223062992, + 0.01590202935039997, + 0.01274480577558279, + 0.015368674881756306, + 0.010626290924847126, + 0.0034955432638525963, + 0.008361482992768288, + 0.013715928420424461, + 0.009041768498718739, + 0.006437363103032112, + 0.015481782145798206, + 0.019351137802004814, + 0.0197389367967844, + 0.00846361368894577, + 0.0074689313769340515, + 0.01915060542523861, + 0.010851479135453701, + 0.009664211422204971, + 0.008258240297436714, + 0.00866799708455801, + 0.011641762219369411, + 0.004789650905877352, + 0.0023206572514027357, + 0.007518627215176821, + 0.0072044446133077145, + 0.005144526716321707, + 0.007226882968097925, + 0.00690957298502326, + 0.010503610596060753, + 0.0033441714476794004, + 0.004927473608404398, + 0.003939449321478605, + 0.0051292674615979195, + 0.003116562496870756, + 0.017405636608600616, + 0.025461645796895027, + 0.01997675932943821, + 0.008001663722097874, + 0.008636639453470707, + 0.026518497616052628, + 0.012375307269394398, + 0.005485838279128075, + 0.009279078803956509, + 0.011781925335526466, + 0.007657174486666918, + 0.004347318783402443, + 0.010550927370786667, + 0.014841508120298386, + 0.00589343998581171, + 0.004220586270093918, + 0.027772659435868263, + 0.016854746267199516, + 0.025775499641895294, + 0.014169436879456043, + 0.024579988792538643, + 0.019781481474637985, + 0.021076014265418053, + 0.033201709389686584, + 0.007895111106336117, + 0.005140342283993959, + 0.005015063565224409, + 0.0038673304952681065, + 0.004699308425188065, + 0.004691922105848789, + 0.007819779217243195, + 0.0033770124427974224, + 0.01644076220691204, + 0.01512995921075344, + 0.014509559608995914, + 0.010507078841328621, + 0.022287748754024506, + 0.014036910608410835, + 0.016639037057757378, + 0.01271901000291109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.001347706071101129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0566580556333065, + 0.06034787744283676, + 0.21768662333488464, + 0.09792044013738632, + 0.09498871117830276, + 0.1536332219839096, + 0.05387762933969498, + 0.14055901765823364, + 0.01834765262901783, + 0.10598073899745941 + ], + "credit_map": [ + 0.008196261711418629, + 0.0069578965194523335, + 0.009353429079055786, + 0.00613251980394125, + 0.005701328627765179, + 0.006440029013901949, + 0.008755732327699661, + 0.005120863206684589, + 0.0070671685971319675, + 0.009046487510204315, + 0.013575068674981594, + 0.0068712541833519936, + 0.005767920520156622, + 0.006312765181064606, + 0.0064375558868050575, + 0.005269656423479319, + 0.026739850640296936, + 0.05422411486506462, + 0.02161616086959839, + 0.01131997350603342, + 0.01281679980456829, + 0.06615548580884933, + 0.01505936961621046, + 0.009754875674843788, + 0.014253509230911732, + 0.014208376407623291, + 0.01383204571902752, + 0.009843781590461731, + 0.010035743936896324, + 0.011880375444889069, + 0.015032089315354824, + 0.00883451383560896, + 0.013103913515806198, + 0.012954581528902054, + 0.016066916286945343, + 0.009098830632865429, + 0.008539768867194653, + 0.011366148479282856, + 0.013709934428334236, + 0.010148616507649422, + 0.012892327271401882, + 0.025225259363651276, + 0.03475766256451607, + 0.007956520654261112, + 0.01883353292942047, + 0.026383377611637115, + 0.017585190013051033, + 0.009999344125390053, + 0.007110896520316601, + 0.008321885019540787, + 0.008092096075415611, + 0.004613528028130531, + 0.004623414482921362, + 0.011001101695001125, + 0.006438812240958214, + 0.003675899701192975, + 0.021748514845967293, + 0.01631176844239235, + 0.015680287033319473, + 0.010129689238965511, + 0.011199461296200752, + 0.02033473178744316, + 0.015355788171291351, + 0.02979877218604088, + 0.002917510224506259, + 0.0020586338359862566, + 0.0036316614132374525, + 0.0014522955752909184, + 0.001703196787275374, + 0.0019533089362084866, + 0.002374995732679963, + 0.002256050007417798, + 0.018977761268615723, + 0.009526882320642471, + 0.0179273784160614, + 0.0055557736195623875, + 0.012625176459550858, + 0.013826854526996613, + 0.0074243308044970036, + 0.020116575062274933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.008641467429697514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129014641046524 + ], + "opposite_logprob_tokens": [ + -2.237901449203491 + ], + "logprob_reward": -0.1129014641046524, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129014641046524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04912181198596954, + 0.040121614933013916, + 0.18617220222949982, + 0.03635204955935478, + 0.1253102868795395, + 0.0782325491309166, + 0.027224915102124214, + 0.14695225656032562, + 0.020221522077918053, + 0.29029080271720886 + ], + "credit_map": [ + 0.007806344423443079, + 0.006236519198864698, + 0.010964136570692062, + 0.004554120358079672, + 0.004440261982381344, + 0.005257965996861458, + 0.0041056787595152855, + 0.005756784230470657, + 0.005271469242870808, + 0.006252503953874111, + 0.006410855799913406, + 0.002662148093804717, + 0.004496782552450895, + 0.007924869656562805, + 0.002647712826728821, + 0.004455276764929295, + 0.02093106508255005, + 0.053166087716817856, + 0.016846001148223877, + 0.00785039458423853, + 0.008494261652231216, + 0.05921609699726105, + 0.011081569828093052, + 0.00858671497553587, + 0.00532856211066246, + 0.004863472655415535, + 0.006571833044290543, + 0.003586658276617527, + 0.003892764914780855, + 0.004376164171844721, + 0.004060403909534216, + 0.003672192106023431, + 0.016173310577869415, + 0.01923179067671299, + 0.028675569221377373, + 0.009246133267879486, + 0.012367057614028454, + 0.01178168598562479, + 0.01434727106243372, + 0.013487462885677814, + 0.010188116692006588, + 0.015967950224876404, + 0.015917347744107246, + 0.0043496498838067055, + 0.006664603482931852, + 0.014554479159414768, + 0.006320865359157324, + 0.004269538912922144, + 0.005930379964411259, + 0.004196640569716692, + 0.002989771543070674, + 0.0014263088814914227, + 0.0023381162900477648, + 0.004953380674123764, + 0.0032689725048840046, + 0.0021213460713624954, + 0.028925905004143715, + 0.01321825385093689, + 0.020071426406502724, + 0.01074331346899271, + 0.027098949998617172, + 0.01258333120495081, + 0.017197178676724434, + 0.017113905400037766, + 0.0023916440550237894, + 0.0035857497714459896, + 0.0028345442842692137, + 0.0016887347446754575, + 0.001968699973076582, + 0.003069188678637147, + 0.0028348946943879128, + 0.00184806645847857, + 0.054624076932668686, + 0.02089153788983822, + 0.0501636266708374, + 0.01302315853536129, + 0.02271845005452633, + 0.03569407761096954, + 0.02043243497610092, + 0.07274343073368073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.1129014641046524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017145470483228564 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017145470483228564, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017145470483228564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07131851464509964, + 0.0676579624414444, + 0.1039743423461914, + 0.07581249624490738, + 0.16456393897533417, + 0.28198233246803284, + 0.03959130123257637, + 0.08364228904247284, + 0.02972111850976944, + 0.08173578232526779 + ], + "credit_map": [ + 0.014267946593463421, + 0.010423589497804642, + 0.013248934410512447, + 0.005713102873414755, + 0.006849127355962992, + 0.01101930346339941, + 0.00547068752348423, + 0.004325817339122295, + 0.014366615563631058, + 0.011827260255813599, + 0.013805204071104527, + 0.004018282983452082, + 0.004077738616615534, + 0.011103622615337372, + 0.0047256010584533215, + 0.0037336291279643774, + 0.014691569842398167, + 0.024418599903583527, + 0.010637719184160233, + 0.005360730923712254, + 0.005293913185596466, + 0.03228353336453438, + 0.004788135644048452, + 0.006500132381916046, + 0.013176650740206242, + 0.010143764317035675, + 0.01548432745039463, + 0.007105575408786535, + 0.0048611704260110855, + 0.010944745503365993, + 0.008540530689060688, + 0.005555722862482071, + 0.030553452670574188, + 0.025778068229556084, + 0.038094278424978256, + 0.011158573441207409, + 0.01712963357567787, + 0.012970242649316788, + 0.017661286517977715, + 0.01121837180107832, + 0.0406690314412117, + 0.05471218377351761, + 0.0672929435968399, + 0.018761444836854935, + 0.02534228004515171, + 0.03482843562960625, + 0.024339552968740463, + 0.01603645086288452, + 0.0053591663017869, + 0.008522887714207172, + 0.0065750787034630775, + 0.0029447844717651606, + 0.0027515962719917297, + 0.005903190467506647, + 0.004522929433733225, + 0.003011659486219287, + 0.0169155765324831, + 0.010198632255196571, + 0.011568581685423851, + 0.006270220503211021, + 0.015926290303468704, + 0.00764322467148304, + 0.009526509791612625, + 0.005593240261077881, + 0.005816522985696793, + 0.00355461984872818, + 0.004161061719059944, + 0.002212615218013525, + 0.0034256670624017715, + 0.0036502573639154434, + 0.0034304731525480747, + 0.003469897201284766, + 0.013964282348752022, + 0.00687469681724906, + 0.011009208858013153, + 0.005954249296337366, + 0.009050601162016392, + 0.009831713512539864, + 0.012365608476102352, + 0.012685414403676987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.0017145470483228564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic long shot sweeps across a vast, sun-drenched meadow in a slow, graceful arc. At the center of this idyllic scene, a woman with long, flowing auburn hair twirls effortlessly. She wears a vibrant, long, earth-yellow dress crafted from delicate, lightweight chiffon that catches the breeze. As she steadily spins on the soles of her feet, her arms extended for balance, her head tilted back in blissful abandon, a serene smile on her face. The vibrant yellow fabric of her dress forms a perfect circle around her, creating a striking visual contrast with the lush green grass dotted with small wildflowers. The entire scene is bathed in the warm golden glow of the afternoon sun, which, as the camera pans, casts long, soft shadows and creates cinematic lens flares, backlighting her body and making the translucent material of her dress shimmer. The atmosphere is one of pure, unadulterated joy and liberated tranquility, capturing a fleeting, dreamlike moment of perfect freedom.", + "seed": 20812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_4/index_5021_seed_20812_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024316433153580876, + "vlm_logprob": -0.024316433153580876, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.024316433153580876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03234865888953209 + ], + "opposite_logprob_tokens": [ + -3.5323486328125 + ], + "logprob_reward": -0.03234865888953209, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03234865888953209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3994118571281433, + 0.2859794497489929, + 0.045393723994493484, + 0.022651223465800285, + 0.050673000514507294, + 0.025221582502126694, + 0.013104853220283985, + 0.036047011613845825, + 0.012027508579194546, + 0.10948978364467621 + ], + "credit_map": [ + 0.04771895706653595, + 0.027819799259305, + 0.09437836706638336, + 0.029367102310061455, + 0.01690247654914856, + 0.12315736711025238, + 0.030392643064260483, + 0.029675129801034927, + 0.06897594034671783, + 0.01512431912124157, + 0.01641920767724514, + 0.01096788328140974, + 0.012189571745693684, + 0.0238337442278862, + 0.12020055949687958, + 0.018268201500177383, + 0.00727895786985755, + 0.004269236233085394, + 0.01241202000528574, + 0.003940302412956953, + 0.0033805829007178545, + 0.0053369225934147835, + 0.0035726940259337425, + 0.00520301004871726, + 0.00303902686573565, + 0.0021967380307614803, + 0.0027811178006231785, + 0.0015101851895451546, + 0.0013964528916403651, + 0.007501828484237194, + 0.001890529878437519, + 0.0023353444412350655, + 0.006907632574439049, + 0.0033060817513614893, + 0.006356185767799616, + 0.0051740966737270355, + 0.00435451976954937, + 0.004376769997179508, + 0.01300464291125536, + 0.007193073630332947, + 0.003505799686536193, + 0.0014599052956327796, + 0.0030781696550548077, + 0.0024551262613385916, + 0.002981272991746664, + 0.003614530200138688, + 0.004034017212688923, + 0.004092761781066656, + 0.0018136975122615695, + 0.001266112201847136, + 0.0029401802457869053, + 0.0009235990582965314, + 0.0014326050877571106, + 0.0013306712498888373, + 0.001447730464860797, + 0.0019502561772242188, + 0.0045891255140304565, + 0.002375248819589615, + 0.00894407369196415, + 0.0020973256323486567, + 0.002494153333827853, + 0.0046531702391803265, + 0.002372940769419074, + 0.008520975708961487, + 0.0011640251614153385, + 0.0007823716732673347, + 0.002174578607082367, + 0.0018063978059217334, + 0.0014390654396265745, + 0.0013259450206533074, + 0.0015895883552730083, + 0.0017455366905778646, + 0.009034830145537853, + 0.0029823475051671267, + 0.03954092040657997, + 0.004716223571449518, + 0.0051064258441329, + 0.03375553712248802, + 0.004906837362796068, + 0.009446649812161922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03234865888953209 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009328543092124164 + ], + "opposite_logprob_tokens": [ + -7.000932693481445 + ], + "logprob_reward": -0.0009328543092124164, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009328543092124164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19076912105083466, + 0.6362329721450806, + 0.051906127482652664, + 0.015418617986142635, + 0.05353100597858429, + 0.009617786854505539, + 0.013312404975295067, + 0.005895492620766163, + 0.0051775239408016205, + 0.018138933926820755 + ], + "credit_map": [ + 0.03462304174900055, + 0.011966871097683907, + 0.031506266444921494, + 0.009796898812055588, + 0.0085086515173316, + 0.06875979900360107, + 0.014220649376511574, + 0.011386938393115997, + 0.16981764137744904, + 0.027971254661679268, + 0.030540233477950096, + 0.024640467017889023, + 0.025565849617123604, + 0.04554115980863571, + 0.2692996561527252, + 0.04285665974020958, + 0.0040267640724778175, + 0.0034566796384751797, + 0.016798585653305054, + 0.0028887952212244272, + 0.0019846917130053043, + 0.01580745354294777, + 0.004151911474764347, + 0.0027912443038076162, + 0.0029079881496727467, + 0.001074424246326089, + 0.001286649378016591, + 0.0008515819790773094, + 0.0008447043946944177, + 0.006071820855140686, + 0.0011182216694578528, + 0.0012632278958335519, + 0.006723650265485048, + 0.0035284897312521935, + 0.005624089855700731, + 0.005880736745893955, + 0.003521418897435069, + 0.005115846171975136, + 0.014740269631147385, + 0.008396507240831852, + 0.001295777503401041, + 0.0006511003593914211, + 0.0010695528471842408, + 0.001246484462171793, + 0.0010727591579779983, + 0.001166756497696042, + 0.0014650353696197271, + 0.0016503204824402928, + 0.0024661256466060877, + 0.0008530016057193279, + 0.0029700426384806633, + 0.001105075585655868, + 0.0008984339074231684, + 0.0022143982350826263, + 0.001333578024059534, + 0.0014717502053827047, + 0.0007601699908263981, + 0.00038122080150060356, + 0.0011441450333222747, + 0.0006709895678795874, + 0.0005882291006855667, + 0.0006747425068169832, + 0.0007348639192059636, + 0.0009411318460479379, + 0.0005095734959468246, + 0.00030003717984072864, + 0.00097199238371104, + 0.0006130033871158957, + 0.00048519132542423904, + 0.001138431136496365, + 0.0006425173487514257, + 0.0005167776253074408, + 0.0014776157913729548, + 0.0005085892044007778, + 0.0063799526542425156, + 0.0008842557435855269, + 0.0008412376628257334, + 0.005847345571964979, + 0.0008070392068475485, + 0.0013928971020504832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0009328543092124164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058646369725465775, + 0.7179262638092041, + 0.03243197500705719, + 0.0498611219227314, + 0.02829308807849884, + 0.021325016394257545, + 0.016036147251725197, + 0.007273093797266483, + 0.02980799227952957, + 0.038398921489715576 + ], + "credit_map": [ + 0.01296866312623024, + 0.004743305034935474, + 0.011603008024394512, + 0.004693241789937019, + 0.003764387685805559, + 0.008833585307002068, + 0.006188057363033295, + 0.005852121859788895, + 0.1896173655986786, + 0.03132963180541992, + 0.03520764783024788, + 0.027153989300131798, + 0.028694089502096176, + 0.05301682651042938, + 0.30478331446647644, + 0.04812336713075638, + 0.0028699440881609917, + 0.0023523964919149876, + 0.00805792585015297, + 0.002826739801093936, + 0.002860161941498518, + 0.004759841598570347, + 0.006242900155484676, + 0.0024620681069791317, + 0.005151663441210985, + 0.002094106748700142, + 0.01103777065873146, + 0.0025880816392600536, + 0.002499172231182456, + 0.02020399086177349, + 0.00345013290643692, + 0.002836202969774604, + 0.0036923049483448267, + 0.0016022276831790805, + 0.0033522529993206263, + 0.0028841334860771894, + 0.0019256176892668009, + 0.003262078622356057, + 0.00661084521561861, + 0.004963630344718695, + 0.0024797474034130573, + 0.0013270957861095667, + 0.0022571233566850424, + 0.0016780232544988394, + 0.002585417591035366, + 0.003618210321292281, + 0.003580729942768812, + 0.003798669669777155, + 0.0016185707645490766, + 0.0011206782655790448, + 0.0018638044130057096, + 0.0018280647927895188, + 0.0026972556952387094, + 0.0018200830090790987, + 0.003477670019492507, + 0.0016100205248221755, + 0.0008855377673171461, + 0.0005323448567651212, + 0.0010822754120454192, + 0.0009658998460508883, + 0.00078107527224347, + 0.0008252901025116444, + 0.0013023980427533388, + 0.0008982720319181681, + 0.003046932630240917, + 0.0015798693057149649, + 0.008514043875038624, + 0.0017310623079538345, + 0.002025030320510268, + 0.008008813485503197, + 0.0022464594803750515, + 0.0026557815726846457, + 0.003105917014181614, + 0.0009889253415167332, + 0.014073501341044903, + 0.0019640233367681503, + 0.0016702985158190131, + 0.012131240218877792, + 0.0015733774052932858, + 0.002891638781875372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895419001579285 + ], + "opposite_logprob_tokens": [ + -2.5789542198181152 + ], + "logprob_reward": -0.07895419001579285, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895419001579285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18393279612064362, + 0.2802434265613556, + 0.10078529268503189, + 0.14557631313800812, + 0.05102989450097084, + 0.024752220138907433, + 0.0412495918571949, + 0.01611383818089962, + 0.09477400779724121, + 0.06154262274503708 + ], + "credit_map": [ + 0.03349289298057556, + 0.01551756914705038, + 0.014925277791917324, + 0.009732595644891262, + 0.008948917500674725, + 0.06381931900978088, + 0.01347248163074255, + 0.024023758247494698, + 0.07920852303504944, + 0.01343624945729971, + 0.014938877895474434, + 0.011466418392956257, + 0.012444798834621906, + 0.015527133829891682, + 0.11451741307973862, + 0.018704034388065338, + 0.009389872662723064, + 0.008208636194467545, + 0.03685189038515091, + 0.006606434006243944, + 0.004611417651176453, + 0.020761843770742416, + 0.007773562800139189, + 0.006581646390259266, + 0.02686200477182865, + 0.006988758221268654, + 0.01773875579237938, + 0.004842379130423069, + 0.006450951565057039, + 0.06535978615283966, + 0.008497253060340881, + 0.008836435154080391, + 0.008006607182323933, + 0.003096923464909196, + 0.008336972445249557, + 0.004774364177137613, + 0.0039468626491725445, + 0.007537671830505133, + 0.008691401220858097, + 0.006639094557613134, + 0.0028458074666559696, + 0.0015553260454908013, + 0.0032021012157201767, + 0.002536144107580185, + 0.002964228857308626, + 0.003602158511057496, + 0.004812130704522133, + 0.0032343235798180103, + 0.006931074429303408, + 0.0023482658434659243, + 0.009167302399873734, + 0.003928222227841616, + 0.004151063039898872, + 0.0054941000416874886, + 0.005344748497009277, + 0.003884816775098443, + 0.002108030952513218, + 0.0010369550436735153, + 0.0032766961958259344, + 0.0012203174410387874, + 0.0019642722327262163, + 0.0023367127869278193, + 0.002003525383770466, + 0.0021673287265002728, + 0.010649754665791988, + 0.004977005068212748, + 0.02793165296316147, + 0.00512586859986186, + 0.004547106102108955, + 0.026460828259587288, + 0.005648085381835699, + 0.009433716535568237, + 0.005426690448075533, + 0.0020749326795339584, + 0.021194541826844215, + 0.0029456859920173883, + 0.0034593380987644196, + 0.018458319827914238, + 0.0034158334601670504, + 0.0045672873966395855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.07895419001579285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a vast, sun-drenched meadow with lush green grass dotted with small wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with long, flowing auburn hair wearing a vibrant, long, earth-yellow chiffon dress visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman spinning on the soles of her feet with her arms extended and head tilted back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, graceful arc (long shot sweep) across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by the sun, creating cinematic lens flares and making the translucent dress shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017159751150757074 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017159751150757074, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017159751150757074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1403471827507019, + 0.5771701335906982, + 0.06508723646402359, + 0.054193876683712006, + 0.07674743980169296, + 0.03138543665409088, + 0.01215222105383873, + 0.008390878327190876, + 0.006835653446614742, + 0.027689943090081215 + ], + "credit_map": [ + 0.01723141223192215, + 0.009347706101834774, + 0.03228780999779701, + 0.01017939206212759, + 0.005797687452286482, + 0.04687563329935074, + 0.010265493765473366, + 0.008362038992345333, + 0.14857235550880432, + 0.026275070384144783, + 0.02608698047697544, + 0.02221963368356228, + 0.023054219782352448, + 0.04832952469587326, + 0.24426141381263733, + 0.03837088495492935, + 0.006308513227850199, + 0.004787299316376448, + 0.02340642921626568, + 0.004502152558416128, + 0.0030064366292208433, + 0.009704210795462132, + 0.008999506942927837, + 0.004372690338641405, + 0.0066826073452830315, + 0.0035637712571769953, + 0.020431291311979294, + 0.00245844479650259, + 0.002111670561134815, + 0.014201641082763672, + 0.0023127899039536715, + 0.0024316650815308094, + 0.009752442128956318, + 0.0044919331558048725, + 0.008794069290161133, + 0.008594570681452751, + 0.005157218314707279, + 0.0070884861052036285, + 0.021254532039165497, + 0.011614185757935047, + 0.003847801359370351, + 0.0019981074146926403, + 0.004100641701370478, + 0.0032075040508061647, + 0.0038855555467307568, + 0.0037042852491140366, + 0.005708689335733652, + 0.004932851064950228, + 0.0022058640606701374, + 0.0009622009820304811, + 0.0025598290376365185, + 0.000980784883722663, + 0.0010672500357031822, + 0.002093392424285412, + 0.0012141629122197628, + 0.0010687375906854868, + 0.0012909190263599157, + 0.0005629842635244131, + 0.0017693248810246587, + 0.000765252101700753, + 0.0007806762005202472, + 0.0010046912357211113, + 0.0009541403269395232, + 0.0012628906406462193, + 0.0007973461179062724, + 0.000405772210797295, + 0.0015148260863497853, + 0.0006096320576034486, + 0.0005793627351522446, + 0.0013264487497508526, + 0.0006218354101292789, + 0.0009804295841604471, + 0.0022919466719031334, + 0.0007341619348153472, + 0.009941195137798786, + 0.0011192637030035257, + 0.001158295082859695, + 0.00925835594534874, + 0.0009763923590071499, + 0.0022103318478912115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0017159751150757074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_tensor([5626, 5626])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_tensor([5626, 5626])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b73babcb8f1f8e71fecc746ca9e2973c111374ce --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_tensor([5626, 5626])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shot from a level angle, boasts a symmetrical and stable composition, focusing on a young monk with a shaved head, sitting cross-legged in the perfect lotus position in the center of the frame. He wears a simple orange-red cotton and linen robe, loosely wrapped around his slender figure, its natural folds creating a rich, layered effect around his legs. His delicate face is slightly closed, his long eyelashes casting a subtle shadow on his lower lids. A calm and focused expression transcends his years, his brows etched with a calming focus. His hands rest in front of him in a meditation mudra. He sits on the dark brown wooden floor of an ancient Zen room, framed by an antique wooden lattice window. A soft beam of afternoon sunlight filters through the panes, casting dappled shadows on the floor and framing the side of his body, creating a warm golden edge. His body was almost motionless, only his chest was rising and falling imperceptibly with his slow, deep breathing. A wisp of green smoke rose from the incense burner not far away and slowly swirled in the still air. The whole scene, rendered in warm colors, was filled with an extremely tranquil, solemn and Zen-like introspective atmosphere.", + "seed": 7593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_5626_seed_7593_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019105827901512384, + "vlm_logprob": -0.019105827901512384, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.019105827901512384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014614085666835308 + ], + "opposite_logprob_tokens": [ + -4.764614105224609 + ], + "logprob_reward": -0.014614085666835308, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014614085666835308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8528924584388733, + 0.036541897803545, + 0.014799905009567738, + 0.013962400145828724, + 0.008333016186952591, + 0.01989748701453209, + 0.016586005687713623, + 0.008925126865506172, + 0.017135515809059143, + 0.010926083661615849 + ], + "credit_map": [ + 0.03283240273594856, + 0.015652989968657494, + 0.34021854400634766, + 0.03906254470348358, + 0.029732905328273773, + 0.026448870077729225, + 0.04362604394555092, + 0.32531824707984924, + 0.0025101464707404375, + 0.003817227203398943, + 0.006445682141929865, + 0.0034648797009140253, + 0.004430166911333799, + 0.002623685635626316, + 0.006655953358858824, + 0.006594161503016949, + 0.0011435144115239382, + 0.0015577448066323996, + 0.0018399429973214865, + 0.001547065214253962, + 0.0028985843528062105, + 0.0017009936273097992, + 0.0025515325833112, + 0.0015605281805619597, + 0.0009812458883970976, + 0.0008399089565500617, + 0.0018645164091140032, + 0.0016010511899366975, + 0.002620273269712925, + 0.001245160587131977, + 0.00323357037268579, + 0.0015766749856993556, + 0.000609643233474344, + 0.0004894720041193068, + 0.0013323428574949503, + 0.00098650180734694, + 0.0010591454338282347, + 0.0010196607327088714, + 0.0016960569191724062, + 0.001140193548053503, + 0.0009450515499338508, + 0.0012320253299549222, + 0.0037851205561310053, + 0.0020056096836924553, + 0.001967458054423332, + 0.0016503765946254134, + 0.003752791555598378, + 0.004559054970741272, + 0.0010520737851038575, + 0.001130486372858286, + 0.002177430083975196, + 0.0019606854766607285, + 0.002326197922229767, + 0.001260145683772862, + 0.0044865780510008335, + 0.0021924099419265985, + 0.0006474521360360086, + 0.0004996879142709076, + 0.0013857305748388171, + 0.0009175099548883736, + 0.0012895037652924657, + 0.0011102486168965697, + 0.002086120657622814, + 0.0009888742351904511, + 0.001189705915749073, + 0.0012137150624766946, + 0.0023482812102884054, + 0.001909421756863594, + 0.0027396169025450945, + 0.001467826426960528, + 0.004075538367033005, + 0.0021914117969572544, + 0.0006734184571541846, + 0.0007206694572232664, + 0.0016478480538353324, + 0.001678560278378427, + 0.0015110812382772565, + 0.0009657650953158736, + 0.002504822099581361, + 0.0012239203788340092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014614085666835308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639241870492697 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639241870492697, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639241870492697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7852416634559631, + 0.03823596239089966, + 0.025775860995054245, + 0.020684864372015, + 0.028758564963936806, + 0.027471374720335007, + 0.013187956064939499, + 0.03240184858441353, + 0.019129792228341103, + 0.009112061932682991 + ], + "credit_map": [ + 0.03072955273091793, + 0.014321276918053627, + 0.3120710253715515, + 0.0354892872273922, + 0.027444176375865936, + 0.02517516352236271, + 0.04134462773799896, + 0.2986666262149811, + 0.0024494619574397802, + 0.003859371179714799, + 0.005211638752371073, + 0.0035732295364141464, + 0.004802393261343241, + 0.00222300598397851, + 0.009038816206157207, + 0.007078045047819614, + 0.00166375283151865, + 0.0021682865917682648, + 0.004204262048006058, + 0.0033871696796268225, + 0.0033773949835449457, + 0.0021937210112810135, + 0.005384062882512808, + 0.0033972107339650393, + 0.0015927973436191678, + 0.0016086812829598784, + 0.003210497321560979, + 0.002468943828716874, + 0.003485815366730094, + 0.0018548035295680165, + 0.004122219048440456, + 0.002341107465326786, + 0.0019787715282291174, + 0.002570487791672349, + 0.004237165674567223, + 0.003958654124289751, + 0.003564194543287158, + 0.002799997804686427, + 0.0061545418575406075, + 0.0034947500098496675, + 0.001650183112360537, + 0.0016513100126758218, + 0.00407829275354743, + 0.0035721592139452696, + 0.0031767813488841057, + 0.002654066774994135, + 0.006436945870518684, + 0.004251635167747736, + 0.0011131714563816786, + 0.0010497565381228924, + 0.002060474595054984, + 0.0012773234630003572, + 0.0018693509045988321, + 0.0012705279514193535, + 0.002807577373459935, + 0.0017397735500708222, + 0.0022099169436842203, + 0.002261348534375429, + 0.004612340591847897, + 0.004472010303288698, + 0.005209712777286768, + 0.0029229072388261557, + 0.0067214625887572765, + 0.003992145881056786, + 0.0013473120052367449, + 0.001179862068966031, + 0.0027217036113142967, + 0.003079250454902649, + 0.0022753654047846794, + 0.0015352688496932387, + 0.003992138896137476, + 0.002998890820890665, + 0.0006951717659831047, + 0.0006915194098837674, + 0.0013569282600656152, + 0.0014547454193234444, + 0.0011152956867590547, + 0.0009703774121589959, + 0.001661083661019802, + 0.0011669403174892068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007639241870492697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059445157647133 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.02059445157647133, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059445157647133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7765569686889648, + 0.043429307639598846, + 0.04030468687415123, + 0.02057318575680256, + 0.020321007817983627, + 0.01599113456904888, + 0.022632300853729248, + 0.02854401245713234, + 0.01557127945125103, + 0.01607605628669262 + ], + "credit_map": [ + 0.032864660024642944, + 0.01893567107617855, + 0.3106343746185303, + 0.034808721393346786, + 0.02929200604557991, + 0.028015751391649246, + 0.030595386400818825, + 0.2914104759693146, + 0.0035118614323437214, + 0.003806335385888815, + 0.00726994639262557, + 0.004421337973326445, + 0.0041814446449279785, + 0.002682085381820798, + 0.008487311191856861, + 0.009068991057574749, + 0.0020951959304511547, + 0.002518628491088748, + 0.006510947830975056, + 0.0051271747797727585, + 0.005682595539838076, + 0.0016791175585240126, + 0.01151373703032732, + 0.0051772939041256905, + 0.0018106260104104877, + 0.0012810094049200416, + 0.0029021366499364376, + 0.0024394176434725523, + 0.003240035381168127, + 0.0023367153480648994, + 0.003486606292426586, + 0.00307664112187922, + 0.0013268996262922883, + 0.002023789333179593, + 0.0033445791341364384, + 0.002493531908839941, + 0.0021731869783252478, + 0.0020395873580127954, + 0.004127253778278828, + 0.002792182844132185, + 0.0012697821948677301, + 0.0008456646464765072, + 0.002565086353570223, + 0.0017403820529580116, + 0.0016054949956014752, + 0.001673077349551022, + 0.003749287221580744, + 0.0025423625484108925, + 0.0016414022538810968, + 0.0019090707646682858, + 0.003455463098362088, + 0.0020343682263046503, + 0.0032896532211452723, + 0.0018146284855902195, + 0.005111507140100002, + 0.003376210108399391, + 0.0025519186165183783, + 0.002764497883617878, + 0.004274341277778149, + 0.002416214207187295, + 0.00513972993940115, + 0.003842045785859227, + 0.003931301645934582, + 0.003623965661972761, + 0.0010042497888207436, + 0.001090640900656581, + 0.0028813532553613186, + 0.0021465467289090157, + 0.0019491312559694052, + 0.0009591945563443005, + 0.003418826963752508, + 0.002121337689459324, + 0.0011679178569465876, + 0.0011823171516880393, + 0.002057470614090562, + 0.0019297568360343575, + 0.0031139268539845943, + 0.0011733074206858873, + 0.003620471339672804, + 0.0018308893777430058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02059445157647133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343872770667076 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343872770667076, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343872770667076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41138792037963867, + 0.08994933217763901, + 0.06536421924829483, + 0.060187119990587234, + 0.12803135812282562, + 0.10005681216716766, + 0.03418165072798729, + 0.03839820995926857, + 0.0417446605861187, + 0.030698686838150024 + ], + "credit_map": [ + 0.01434856466948986, + 0.00850444845855236, + 0.15458352863788605, + 0.020295387133955956, + 0.01621328666806221, + 0.01445747446268797, + 0.028823453933000565, + 0.15416176617145538, + 0.005652694497257471, + 0.005920207127928734, + 0.013623975217342377, + 0.011207184754312038, + 0.012478826567530632, + 0.006221395917236805, + 0.02320953644812107, + 0.011635511182248592, + 0.0036876793019473553, + 0.004098890349268913, + 0.010632030665874481, + 0.008804001845419407, + 0.008609901182353497, + 0.004576509818434715, + 0.01597917266190052, + 0.008976036682724953, + 0.004666710272431374, + 0.004895086400210857, + 0.007746222894638777, + 0.005885607562959194, + 0.010885117575526237, + 0.006085067987442017, + 0.012719360180199146, + 0.007303946651518345, + 0.007760871667414904, + 0.010240398347377777, + 0.01812073029577732, + 0.016909323632717133, + 0.016950905323028564, + 0.010426141321659088, + 0.030171262100338936, + 0.017451724037528038, + 0.006205704528838396, + 0.005945437587797642, + 0.014687258750200272, + 0.013982499949634075, + 0.011745527386665344, + 0.008223687298595905, + 0.023433104157447815, + 0.015833592042326927, + 0.0025345187168568373, + 0.0023843469098210335, + 0.005058246664702892, + 0.003400593064725399, + 0.003692593425512314, + 0.0029226995538920164, + 0.009147834032773972, + 0.005040820222347975, + 0.002951557282358408, + 0.0025559004861861467, + 0.005829055327922106, + 0.004390796180814505, + 0.0056618815287947655, + 0.003774502547457814, + 0.007529742084443569, + 0.005704777780920267, + 0.0018875983078032732, + 0.002386799082159996, + 0.006961051374673843, + 0.0063354321755468845, + 0.004440258257091045, + 0.0018905634060502052, + 0.011681588366627693, + 0.0061613693833351135, + 0.0016136392951011658, + 0.0015565349021926522, + 0.004225977696478367, + 0.0034507568925619125, + 0.004062269348651171, + 0.002086061052978039, + 0.009583820588886738, + 0.0041196285746991634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.026343872770667076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337487623095512 + ], + "opposite_logprob_tokens": [ + -3.6513373851776123 + ], + "logprob_reward": -0.026337487623095512, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337487623095512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7063533663749695, + 0.12578879296779633, + 0.025468166917562485, + 0.028195654973387718, + 0.024004854261875153, + 0.011529543437063694, + 0.02509203553199768, + 0.028885291889309883, + 0.014516834169626236, + 0.010165520943701267 + ], + "credit_map": [ + 0.02818235754966736, + 0.015307304449379444, + 0.2856781482696533, + 0.030438875779509544, + 0.023883532732725143, + 0.02371804416179657, + 0.026106009259819984, + 0.2730390429496765, + 0.009231285192072392, + 0.015741413459181786, + 0.02289254404604435, + 0.01053128857165575, + 0.01571750082075596, + 0.014869451522827148, + 0.013338230550289154, + 0.023467060178518295, + 0.002060094615444541, + 0.002246247371658683, + 0.005674168933182955, + 0.0034737822134047747, + 0.0024177846498787403, + 0.002446710132062435, + 0.0034042561892420053, + 0.003745121182873845, + 0.0016723070293664932, + 0.0017021860694512725, + 0.005372410640120506, + 0.003813854418694973, + 0.0032208533957600594, + 0.002189583145081997, + 0.006214350927621126, + 0.0040101055055856705, + 0.0013910458656027913, + 0.0015681248623877764, + 0.003322827396914363, + 0.003585749538615346, + 0.0029627042822539806, + 0.0017715791473165154, + 0.006543282885104418, + 0.002859537024050951, + 0.0007394711719825864, + 0.0006445456529036164, + 0.0023670655209571123, + 0.0008352989098057151, + 0.001335797132924199, + 0.001139939296990633, + 0.001302229124121368, + 0.003165195928886533, + 0.0022443842608481646, + 0.0015969863161444664, + 0.0043746717274188995, + 0.0022753607481718063, + 0.002954938681796193, + 0.0022576688788831234, + 0.004722215700894594, + 0.004665807820856571, + 0.0015093639958649874, + 0.0017076011281460524, + 0.004473340231925249, + 0.003950420767068863, + 0.0038659756537526846, + 0.0016129568684846163, + 0.007609634660184383, + 0.004155994858592749, + 0.0011026390129700303, + 0.000843196117784828, + 0.002394559560343623, + 0.0019014698918908834, + 0.0020784155931323767, + 0.0011657486902549863, + 0.0031947495881468058, + 0.0018360535614192486, + 0.0006369036273099482, + 0.000706795952282846, + 0.0013486910611391068, + 0.0011732177808880806, + 0.0014338173205032945, + 0.0006285074632614851, + 0.002746438141912222, + 0.0014911484904587269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.026337487623095512 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle, boasts a symmetrical and stable composition, focusing on a young monk with a shaved head, sitting cross-legged in the perfect lotus position in the center of the frame. He wears a simple orange-red cotton and linen robe, loosely wrapped around his slender figure, its natural folds creating a rich, layered effect around his legs. His delicate face is slightly closed, his long eyelashes casting a subtle shadow on his lower lids. A calm and focused expression transcends his years, his brows etched with a calming focus. His hands rest in front of him in a meditation mudra. He sits on the dark brown wooden floor of an ancient Zen room, framed by an antique wooden lattice window. A soft beam of afternoon sunlight filters through the panes, casting dappled shadows on the floor and framing the side of his body, creating a warm golden edge. His body was almost motionless, only his chest was rising and falling imperceptibly with his slow, deep breathing. A wisp of green smoke rose from the incense burner not far away and slowly swirled in the still air. The whole scene, rendered in warm colors, was filled with an extremely tranquil, solemn and Zen-like introspective atmosphere.", + "seed": 7594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_5626_seed_7594_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019550530426204204, + "vlm_logprob": -0.019550530426204204, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.019550530426204204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01729384809732437 + ], + "opposite_logprob_tokens": [ + -4.767293930053711 + ], + "logprob_reward": -0.01729384809732437, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01729384809732437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42498934268951416, + 0.08856352418661118, + 0.061107147485017776, + 0.050739988684654236, + 0.04613766819238663, + 0.04856443777680397, + 0.08319485932588577, + 0.06854487955570221, + 0.061728138476610184, + 0.06643004715442657 + ], + "credit_map": [ + 0.0735669732093811, + 0.050609052181243896, + 0.06106093153357506, + 0.0354611799120903, + 0.042128223925828934, + 0.04644562676548958, + 0.06168554723262787, + 0.05403180420398712, + 0.011682989075779915, + 0.009726063348352909, + 0.01292912382632494, + 0.007484397850930691, + 0.007991785183548927, + 0.009585806168615818, + 0.01518225111067295, + 0.013981104828417301, + 0.003442924004048109, + 0.0061464677564799786, + 0.007748913485556841, + 0.0077397022396326065, + 0.00546595873311162, + 0.00913061760365963, + 0.01354236714541912, + 0.007890192791819572, + 0.0065391226671636105, + 0.004628357011824846, + 0.007404376287013292, + 0.003852862399071455, + 0.0038828193210065365, + 0.006196279078722, + 0.01032438687980175, + 0.007911784574389458, + 0.0021654400043189526, + 0.003037817310541868, + 0.007448635529726744, + 0.006791223306208849, + 0.00375235453248024, + 0.002749336650595069, + 0.01421433687210083, + 0.005978525150567293, + 0.0034432781394571066, + 0.0043753101490437984, + 0.005897053051739931, + 0.005150455515831709, + 0.006965115200728178, + 0.004472068976610899, + 0.01029353216290474, + 0.007967625744640827, + 0.005038058385252953, + 0.008972730487585068, + 0.0126081807538867, + 0.008516388013958931, + 0.008585837669670582, + 0.005840788595378399, + 0.022179409861564636, + 0.011453467421233654, + 0.003977426793426275, + 0.006152336020022631, + 0.010452142916619778, + 0.008872206322848797, + 0.005957856774330139, + 0.004610888194292784, + 0.017686109989881516, + 0.010835913009941578, + 0.0052526723593473434, + 0.008598169311881065, + 0.007017824333161116, + 0.006314742844551802, + 0.005953628569841385, + 0.006572554353624582, + 0.012024453841149807, + 0.009994092397391796, + 0.0058281742967665195, + 0.0070442939177155495, + 0.009685622528195381, + 0.009025364182889462, + 0.0051272655837237835, + 0.006578589789569378, + 0.013611527159810066, + 0.00952921062707901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01729384809732437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.424610435962677, + 0.10162342339754105, + 0.0654672235250473, + 0.04812953993678093, + 0.048396699130535126, + 0.04058336466550827, + 0.11395204812288284, + 0.04428707808256149, + 0.05702681466937065, + 0.05592330917716026 + ], + "credit_map": [ + 0.03053874522447586, + 0.036173392087221146, + 0.05753586068749428, + 0.05374493449926376, + 0.050380904227495193, + 0.047042038291692734, + 0.09448458254337311, + 0.05470997095108032, + 0.006228037644177675, + 0.00777249364182353, + 0.015633655712008476, + 0.011499921791255474, + 0.011772369034588337, + 0.00764882005751133, + 0.024632303044199944, + 0.016435818746685982, + 0.004297875799238682, + 0.004173962865024805, + 0.010644987225532532, + 0.009456763043999672, + 0.007712746970355511, + 0.004446233622729778, + 0.015719473361968994, + 0.009015176445245743, + 0.002459314651787281, + 0.0031044615898281336, + 0.008719053119421005, + 0.00616639805957675, + 0.005092773586511612, + 0.0029874190222471952, + 0.012257670052349567, + 0.007342451252043247, + 0.004067264962941408, + 0.002871990902349353, + 0.0073228091932833195, + 0.006566773168742657, + 0.003615773282945156, + 0.0040519204922020435, + 0.011724901385605335, + 0.008175268769264221, + 0.0033765591215342283, + 0.002947768196463585, + 0.005246902350336313, + 0.004635557997971773, + 0.004948883317410946, + 0.004394048359245062, + 0.008935783058404922, + 0.006097861565649509, + 0.00614647613838315, + 0.008649207651615143, + 0.016670208424329758, + 0.012758085504174232, + 0.015241161920130253, + 0.008627869188785553, + 0.02862505055963993, + 0.017233995720744133, + 0.002850915305316448, + 0.002909489441663027, + 0.006898077204823494, + 0.006355974357575178, + 0.0043629612773656845, + 0.0035831057466566563, + 0.010442821308970451, + 0.006883734371513128, + 0.005484888330101967, + 0.006121531128883362, + 0.007803889457136393, + 0.006459565367549658, + 0.006345766130834818, + 0.0059512909501791, + 0.011878199875354767, + 0.006981681566685438, + 0.0031510628759860992, + 0.0038712075911462307, + 0.008682966232299805, + 0.008852644823491573, + 0.005973852705210447, + 0.0038700737059116364, + 0.01436834316700697, + 0.007153154816478491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.009788372553884983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108570583164692 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.01108570583164692, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108570583164692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32796892523765564, + 0.08126357942819595, + 0.059117674827575684, + 0.03963979706168175, + 0.08024782687425613, + 0.09397643804550171, + 0.11272423714399338, + 0.0708484873175621, + 0.06385599821805954, + 0.07035703212022781 + ], + "credit_map": [ + 0.02653515338897705, + 0.03994979336857796, + 0.04339570924639702, + 0.0423850417137146, + 0.031345754861831665, + 0.030223263427615166, + 0.061960384249687195, + 0.052173830568790436, + 0.00432719336822629, + 0.007211998105049133, + 0.014345205388963223, + 0.007316619623452425, + 0.008799209259450436, + 0.006848614197224379, + 0.0182644072920084, + 0.014150330796837807, + 0.003949826117604971, + 0.00457813311368227, + 0.010122072882950306, + 0.006621859967708588, + 0.0057121277786791325, + 0.005464962683618069, + 0.014643343165516853, + 0.008025347255170345, + 0.001967119984328747, + 0.003251760033890605, + 0.0066193994134664536, + 0.004523892886936665, + 0.004588240757584572, + 0.0031439948361366987, + 0.009982193820178509, + 0.0055631957948207855, + 0.007973931729793549, + 0.00574148865416646, + 0.012036040425300598, + 0.008769825100898743, + 0.005789139773696661, + 0.007819360122084618, + 0.018259501084685326, + 0.013858540914952755, + 0.007506603840738535, + 0.007057507522404194, + 0.012329633347690105, + 0.01142718456685543, + 0.01056264154613018, + 0.007719751913100481, + 0.02113853022456169, + 0.016234591603279114, + 0.00792237650603056, + 0.009573735296726227, + 0.015748580917716026, + 0.015098909847438335, + 0.011428220197558403, + 0.007048849016427994, + 0.024929020553827286, + 0.020974548533558846, + 0.006031859200447798, + 0.0054772947914898396, + 0.01163344644010067, + 0.007332991808652878, + 0.0075608957558870316, + 0.007618485018610954, + 0.015938960015773773, + 0.009254558011889458, + 0.0049642170779407024, + 0.005167684983462095, + 0.008226606994867325, + 0.006323510780930519, + 0.008836393244564533, + 0.008148590102791786, + 0.013351180590689182, + 0.008837814442813396, + 0.008460232987999916, + 0.00825978722423315, + 0.007979092188179493, + 0.005210680887103081, + 0.008278866298496723, + 0.010419778525829315, + 0.012985463254153728, + 0.008763129822909832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.01108570583164692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3364236056804657, + 0.12076690047979355, + 0.11522635817527771, + 0.0881481021642685, + 0.06537197530269623, + 0.06656017899513245, + 0.05127226561307907, + 0.06449567526578903, + 0.040059030055999756, + 0.051675885915756226 + ], + "credit_map": [ + 0.02884620614349842, + 0.025028273463249207, + 0.052776437252759933, + 0.03737124800682068, + 0.0343470498919487, + 0.05570944398641586, + 0.06280554831027985, + 0.0395394042134285, + 0.010412594303488731, + 0.008483069017529488, + 0.019701505079865456, + 0.012427926994860172, + 0.011865212582051754, + 0.01704367622733116, + 0.025799613445997238, + 0.015033304691314697, + 0.009874093346297741, + 0.007005246821790934, + 0.0213312990963459, + 0.014718887396156788, + 0.011079934425652027, + 0.009935292415320873, + 0.028261974453926086, + 0.013019626028835773, + 0.00364312925375998, + 0.004816999658942223, + 0.016679754480719566, + 0.011915137059986591, + 0.009592289105057716, + 0.003912277519702911, + 0.02694050781428814, + 0.01064800750464201, + 0.005580424331128597, + 0.003865729784592986, + 0.01159906480461359, + 0.008403949439525604, + 0.006413348019123077, + 0.003942523151636124, + 0.01799161545932293, + 0.0075753191486001015, + 0.0031555425375699997, + 0.005122884176671505, + 0.010953999124467373, + 0.007128363009542227, + 0.007178464904427528, + 0.006127071566879749, + 0.014893298037350178, + 0.012000554241240025, + 0.003668304532766342, + 0.0033369562588632107, + 0.007294694427400827, + 0.006361931562423706, + 0.004554014652967453, + 0.003811748931184411, + 0.011261195875704288, + 0.01098342053592205, + 0.004311933182179928, + 0.005118002183735371, + 0.010052662342786789, + 0.007331066764891148, + 0.0072317905724048615, + 0.005655581597238779, + 0.016041312366724014, + 0.00875332485884428, + 0.003273307578638196, + 0.00289837084710598, + 0.00659392261877656, + 0.004478798247873783, + 0.00417133467271924, + 0.004050181247293949, + 0.009445197880268097, + 0.00514791626483202, + 0.0046865735203027725, + 0.005517679266631603, + 0.007267004810273647, + 0.005758189130574465, + 0.005366090219467878, + 0.005398775450885296, + 0.01022317074239254, + 0.00745840510353446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978975884616375 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.02978975884616375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978975884616375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3469603955745697, + 0.1274193376302719, + 0.03754875808954239, + 0.047753170132637024, + 0.05864628031849861, + 0.09606920182704926, + 0.06024281680583954, + 0.03827909752726555, + 0.05309641361236572, + 0.1339844912290573 + ], + "credit_map": [ + 0.03052830509841442, + 0.03365984931588173, + 0.061205100268125534, + 0.04419654235243797, + 0.03293617442250252, + 0.030286913737654686, + 0.06408814340829849, + 0.05005939304828644, + 0.01747875288128853, + 0.013330565765500069, + 0.02235967479646206, + 0.011664843186736107, + 0.007662282790988684, + 0.015165144577622414, + 0.021218879148364067, + 0.01853920705616474, + 0.003191164927557111, + 0.0043992651626467705, + 0.004972974304109812, + 0.0040135458111763, + 0.0032853996381163597, + 0.004514755215495825, + 0.0066894651390612125, + 0.0064821913838386536, + 0.0057284594513475895, + 0.005059290211647749, + 0.006984760984778404, + 0.0041354927234351635, + 0.0032815372105687857, + 0.0059547279961407185, + 0.00746650155633688, + 0.00914240162819624, + 0.0037678496446460485, + 0.004830031655728817, + 0.008114689029753208, + 0.008412650786340237, + 0.0065782503224909306, + 0.005877222865819931, + 0.011591095477342606, + 0.00947449542582035, + 0.008462067693471909, + 0.012915507890284061, + 0.015943266451358795, + 0.0071562351658940315, + 0.00629472965374589, + 0.01515901181846857, + 0.011676534079015255, + 0.01846185512840748, + 0.008168312720954418, + 0.007266386412084103, + 0.008181201294064522, + 0.006233911961317062, + 0.004759227856993675, + 0.005887982435524464, + 0.009426163509488106, + 0.010319635272026062, + 0.0036138645373284817, + 0.004666284192353487, + 0.005761357024312019, + 0.003378647845238447, + 0.00366751616820693, + 0.004743242636322975, + 0.006109150592237711, + 0.0063390349969267845, + 0.0024623798672109842, + 0.004806295968592167, + 0.0076489076018333435, + 0.005876326002180576, + 0.007078681141138077, + 0.0051516080275177956, + 0.011598031967878342, + 0.008474190719425678, + 0.013030458241701126, + 0.022440269589424133, + 0.018011324107646942, + 0.01536291092634201, + 0.012030907906591892, + 0.015417804010212421, + 0.019806932657957077, + 0.017883896827697754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02978975884616375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle, boasts a symmetrical and stable composition, focusing on a young monk with a shaved head, sitting cross-legged in the perfect lotus position in the center of the frame. He wears a simple orange-red cotton and linen robe, loosely wrapped around his slender figure, its natural folds creating a rich, layered effect around his legs. His delicate face is slightly closed, his long eyelashes casting a subtle shadow on his lower lids. A calm and focused expression transcends his years, his brows etched with a calming focus. His hands rest in front of him in a meditation mudra. He sits on the dark brown wooden floor of an ancient Zen room, framed by an antique wooden lattice window. A soft beam of afternoon sunlight filters through the panes, casting dappled shadows on the floor and framing the side of his body, creating a warm golden edge. His body was almost motionless, only his chest was rising and falling imperceptibly with his slow, deep breathing. A wisp of green smoke rose from the incense burner not far away and slowly swirled in the still air. The whole scene, rendered in warm colors, was filled with an extremely tranquil, solemn and Zen-like introspective atmosphere.", + "seed": 7595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_5626_seed_7595_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021009694784879684, + "vlm_logprob": -0.021009694784879684, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.021009694784879684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020160874351859093 + ], + "opposite_logprob_tokens": [ + -5.145160675048828 + ], + "logprob_reward": -0.020160874351859093, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.020160874351859093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.284947007894516, + 0.06445083767175674, + 0.04338948056101799, + 0.07667860388755798, + 0.14028672873973846, + 0.08456667512655258, + 0.09678177535533905, + 0.09631282836198807, + 0.05252443999052048, + 0.06006160005927086 + ], + "credit_map": [ + 0.032754648476839066, + 0.021375728771090508, + 0.03743002936244011, + 0.03784426301717758, + 0.03642809018492699, + 0.04016238823533058, + 0.045143384486436844, + 0.03380849212408066, + 0.009129103273153305, + 0.009901908226311207, + 0.007604154292494059, + 0.0077610439620912075, + 0.006321045570075512, + 0.012177182361483574, + 0.006038470193743706, + 0.005517938174307346, + 0.0061681317165493965, + 0.002975638024508953, + 0.008557433262467384, + 0.004005059599876404, + 0.004483007825911045, + 0.00630077812820673, + 0.005867093801498413, + 0.005032343789935112, + 0.008811090141534805, + 0.004975223448127508, + 0.013096263632178307, + 0.008729089982807636, + 0.009979144670069218, + 0.012023433111608028, + 0.01038077287375927, + 0.00868359487503767, + 0.01917717419564724, + 0.011267886497080326, + 0.019460778683423996, + 0.014748409390449524, + 0.017439382150769234, + 0.02395956963300705, + 0.015701068565249443, + 0.018532468006014824, + 0.008765134029090405, + 0.004474170506000519, + 0.015863031148910522, + 0.008916188962757587, + 0.00818926002830267, + 0.016761798411607742, + 0.012574061751365662, + 0.00902304332703352, + 0.015028418973088264, + 0.008766194805502892, + 0.011370331048965454, + 0.010578641667962074, + 0.011243704706430435, + 0.01651628315448761, + 0.012944919988512993, + 0.010333294048905373, + 0.015571262687444687, + 0.008071066811680794, + 0.011409375816583633, + 0.009441631846129894, + 0.009835933335125446, + 0.01705106534063816, + 0.01419877540320158, + 0.010733729228377342, + 0.007153199519962072, + 0.003342835232615471, + 0.006465532351285219, + 0.006698383949697018, + 0.007800795137882233, + 0.008919397369027138, + 0.0072378902696073055, + 0.004906408488750458, + 0.008295194245874882, + 0.0058389403857290745, + 0.006530219689011574, + 0.009363136254251003, + 0.005239072721451521, + 0.012119486927986145, + 0.007693272549659014, + 0.004982280544936657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.020160874351859093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108712051063776 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.01108712051063776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108712051063776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.570327639579773, + 0.07999629527330399, + 0.06164894253015518, + 0.036685552448034286, + 0.051842764019966125, + 0.029294118285179138, + 0.06421013176441193, + 0.0310450978577137, + 0.046497657895088196, + 0.028451764956116676 + ], + "credit_map": [ + 0.06216754764318466, + 0.059972088783979416, + 0.0703539103269577, + 0.05697987973690033, + 0.12424109131097794, + 0.09384574741125107, + 0.05684538558125496, + 0.04592199996113777, + 0.008068254217505455, + 0.0060450793243944645, + 0.008184860460460186, + 0.009712628088891506, + 0.018898796290159225, + 0.012711826711893082, + 0.009923264384269714, + 0.0064515890553593636, + 0.0055087595246732235, + 0.004767430014908314, + 0.005701863206923008, + 0.006587401498109102, + 0.01512206345796585, + 0.008219839073717594, + 0.01058866549283266, + 0.005152917932718992, + 0.005116652697324753, + 0.0027363114058971405, + 0.004354387987405062, + 0.0035633703228086233, + 0.0044782282784581184, + 0.008659698069095612, + 0.004577049985527992, + 0.0031998560298234224, + 0.006493529304862022, + 0.004229850135743618, + 0.006158157717436552, + 0.005711518693715334, + 0.007101109717041254, + 0.008325460366904736, + 0.008163140155375004, + 0.005660001654177904, + 0.003716631792485714, + 0.0016625863499939442, + 0.003671244252473116, + 0.00499826529994607, + 0.004595500882714987, + 0.004801172763109207, + 0.00349803757853806, + 0.0023506800644099712, + 0.010754837654531002, + 0.006109622772783041, + 0.009001142345368862, + 0.006358264479786158, + 0.00714625371620059, + 0.009832125157117844, + 0.008066164329648018, + 0.00694171991199255, + 0.003956932574510574, + 0.00162226683460176, + 0.0028422721661627293, + 0.004281348083168268, + 0.0041963281109929085, + 0.004482979886233807, + 0.00533835357055068, + 0.004324616398662329, + 0.00519945565611124, + 0.002635468263179064, + 0.0058390479534864426, + 0.00487250741571188, + 0.005247105378657579, + 0.008533724583685398, + 0.007810920476913452, + 0.006359429564327002, + 0.00393771892413497, + 0.0029226394835859537, + 0.002699225442484021, + 0.0034372727386653423, + 0.0032486622221767902, + 0.0050199320539832115, + 0.003261706791818142, + 0.0039246040396392345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01108712051063776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208580367267132 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208580367267132, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208580367267132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37040603160858154, + 0.08656679093837738, + 0.07121734321117401, + 0.07724529504776001, + 0.05206374451518059, + 0.05389075353741646, + 0.11326362192630768, + 0.05288408324122429, + 0.05864199995994568, + 0.06382036209106445 + ], + "credit_map": [ + 0.05619971826672554, + 0.020164739340543747, + 0.052807413041591644, + 0.05444631725549698, + 0.03437874838709831, + 0.060449060052633286, + 0.05343910679221153, + 0.03852090612053871, + 0.010255170986056328, + 0.008618810214102268, + 0.007436351850628853, + 0.013863290660083294, + 0.011504603549838066, + 0.012203043326735497, + 0.015392474830150604, + 0.007293048780411482, + 0.006718276534229517, + 0.004863436799496412, + 0.0069498335942626, + 0.009608709253370762, + 0.009869802743196487, + 0.010140410624444485, + 0.016451871022582054, + 0.006615004502236843, + 0.01074686273932457, + 0.004929196555167437, + 0.010148181580007076, + 0.0122068515047431, + 0.007664342876523733, + 0.013993433676660061, + 0.009087136015295982, + 0.008469294756650925, + 0.006309809163212776, + 0.002858967985957861, + 0.004485100973397493, + 0.009061151184141636, + 0.008607117459177971, + 0.006374280899763107, + 0.0091975973919034, + 0.005169715732336044, + 0.0072000049985945225, + 0.003482878440991044, + 0.005721254274249077, + 0.00828163418918848, + 0.006471846718341112, + 0.006572658661752939, + 0.010454128496348858, + 0.005706343799829483, + 0.01760094240307808, + 0.010882877744734287, + 0.01758473552763462, + 0.0115693723782897, + 0.012368392199277878, + 0.016412772238254547, + 0.015073255635797977, + 0.011771267279982567, + 0.006189062260091305, + 0.0026653397362679243, + 0.004975565243512392, + 0.00827754195779562, + 0.007205942180007696, + 0.005863760132342577, + 0.010346378199756145, + 0.007360496092587709, + 0.007602402940392494, + 0.005009972490370274, + 0.006150594912469387, + 0.007506238296627998, + 0.005600672215223312, + 0.011755925603210926, + 0.009011064656078815, + 0.0060051302425563335, + 0.00879954919219017, + 0.005021971650421619, + 0.007438614498823881, + 0.008436997421085835, + 0.009504920803010464, + 0.009823650121688843, + 0.008578087203204632, + 0.006216562818735838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.014208580367267132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799478128552437 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799478128552437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799478128552437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2924044132232666, + 0.1252148449420929, + 0.11937632411718369, + 0.07105454802513123, + 0.06741343438625336, + 0.08549606055021286, + 0.07367070764303207, + 0.05413292348384857, + 0.058709293603897095, + 0.05252741649746895 + ], + "credit_map": [ + 0.039234794676303864, + 0.02370423637330532, + 0.04877227172255516, + 0.04180514067411423, + 0.03946874663233757, + 0.03733428567647934, + 0.03603777661919594, + 0.026047157123684883, + 0.010989358648657799, + 0.008677727542817593, + 0.013503133319318295, + 0.01959448680281639, + 0.026929454877972603, + 0.011584760621190071, + 0.025110626593232155, + 0.008825289085507393, + 0.010748730041086674, + 0.010132774710655212, + 0.012604222632944584, + 0.016943229362368584, + 0.024465642869472504, + 0.008438101969659328, + 0.027545372024178505, + 0.008498240262269974, + 0.0067715817131102085, + 0.005651785060763359, + 0.005752788856625557, + 0.011799372732639313, + 0.011767908930778503, + 0.007024600636214018, + 0.015041493810713291, + 0.0072450218722224236, + 0.007253346033394337, + 0.005412928760051727, + 0.005812839604914188, + 0.007868003100156784, + 0.011639081872999668, + 0.011122111231088638, + 0.01118377223610878, + 0.007121350150555372, + 0.008274614810943604, + 0.004973433446139097, + 0.0086818253621459, + 0.010124093852937222, + 0.01742270030081272, + 0.009225302375853062, + 0.018983909860253334, + 0.007810188923031092, + 0.013574049808084965, + 0.005708478856831789, + 0.009505740366876125, + 0.00713524641469121, + 0.008520172908902168, + 0.013423122465610504, + 0.00801781378686428, + 0.007786086294800043, + 0.006363723427057266, + 0.003054666565731168, + 0.0056069353595376015, + 0.0063975197263062, + 0.007479543332010508, + 0.007342474535107613, + 0.012258878909051418, + 0.005629181396216154, + 0.00632481137290597, + 0.0034232663456350565, + 0.006037197075784206, + 0.007166228722780943, + 0.006145378574728966, + 0.010500147007405758, + 0.01199399959295988, + 0.00711826141923666, + 0.007469993084669113, + 0.003810865804553032, + 0.005472242366522551, + 0.006254088599234819, + 0.007930480875074863, + 0.008656605146825314, + 0.008806073106825352, + 0.004127068445086479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029799478128552437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792420566082 + ], + "opposite_logprob_tokens": [ + -3.529792308807373 + ], + "logprob_reward": -0.029792420566082, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792420566082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38892534375190735, + 0.07116073369979858, + 0.08591336011886597, + 0.09077931195497513, + 0.0914936289191246, + 0.11988691985607147, + 0.05563870072364807, + 0.05253858119249344, + 0.028612056747078896, + 0.015051348134875298 + ], + "credit_map": [ + 0.059875842183828354, + 0.024350501596927643, + 0.037812091410160065, + 0.055833544582128525, + 0.04293687269091606, + 0.08050812035799026, + 0.04687166213989258, + 0.04073669761419296, + 0.008376240730285645, + 0.0070362272672355175, + 0.008401158265769482, + 0.00809972733259201, + 0.008107735775411129, + 0.017155174165964127, + 0.00798656977713108, + 0.005997898988425732, + 0.012267286889255047, + 0.009095759131014347, + 0.008465026505291462, + 0.008895127102732658, + 0.010218183510005474, + 0.020184986293315887, + 0.009333034977316856, + 0.007453955244272947, + 0.014693680219352245, + 0.007381245493888855, + 0.01026370283216238, + 0.01109528448432684, + 0.007452134042978287, + 0.021073322743177414, + 0.011141898110508919, + 0.0076780435629189014, + 0.01581961289048195, + 0.005567021202296019, + 0.010923630557954311, + 0.009399163536727428, + 0.008090182207524776, + 0.02132852002978325, + 0.010204989463090897, + 0.010160515084862709, + 0.014779098331928253, + 0.008288849145174026, + 0.013027171604335308, + 0.015869015827775, + 0.012225580401718616, + 0.02643435448408127, + 0.017601650208234787, + 0.011661194264888763, + 0.007323688827455044, + 0.005329948849976063, + 0.0074120559729635715, + 0.005985476542264223, + 0.0065861535258591175, + 0.008182086981832981, + 0.008685735985636711, + 0.006133558228611946, + 0.010512340813875198, + 0.0034288750030100346, + 0.006055560428649187, + 0.005400707479566336, + 0.006132879760116339, + 0.00902294460684061, + 0.006643707398325205, + 0.00534156383946538, + 0.0050908955745399, + 0.0018247788539156318, + 0.003321317723020911, + 0.002806582022458315, + 0.0031863567419350147, + 0.005254025105386972, + 0.004193153697997332, + 0.00293494644574821, + 0.002337329089641571, + 0.0011852189199998975, + 0.0017130844062194228, + 0.0014899231027811766, + 0.0015452289953827858, + 0.0033022763673216105, + 0.0019419953459873796, + 0.0015362921403720975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.029792420566082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a level angle, boasts a symmetrical and stable composition, focusing on a young monk with a shaved head, sitting cross-legged in the perfect lotus position in the center of the frame. He wears a simple orange-red cotton and linen robe, loosely wrapped around his slender figure, its natural folds creating a rich, layered effect around his legs. His delicate face is slightly closed, his long eyelashes casting a subtle shadow on his lower lids. A calm and focused expression transcends his years, his brows etched with a calming focus. His hands rest in front of him in a meditation mudra. He sits on the dark brown wooden floor of an ancient Zen room, framed by an antique wooden lattice window. A soft beam of afternoon sunlight filters through the panes, casting dappled shadows on the floor and framing the side of his body, creating a warm golden edge. His body was almost motionless, only his chest was rising and falling imperceptibly with his slow, deep breathing. A wisp of green smoke rose from the incense burner not far away and slowly swirled in the still air. The whole scene, rendered in warm colors, was filled with an extremely tranquil, solemn and Zen-like introspective atmosphere.", + "seed": 7596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_5/index_5626_seed_7596_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01967358887195587, + "vlm_logprob": -0.01967358887195587, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.01967358887195587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023626580834388733 + ], + "opposite_logprob_tokens": [ + -4.523626804351807 + ], + "logprob_reward": -0.023626580834388733, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.023626580834388733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018858516588807106, + 0.008013429120182991, + 0.038374852389097214, + 0.016803717240691185, + 0.14308835566043854, + 0.02021697163581848, + 0.7113367319107056, + 0.029917676001787186, + 0.006347295828163624, + 0.007042518816888332 + ], + "credit_map": [ + 0.0019306166796013713, + 0.0018938516732305288, + 0.0036538694985210896, + 0.00230632396414876, + 0.0018710024887695909, + 0.0009240476065315306, + 0.002883058739826083, + 0.0033957436680793762, + 0.0008964382577687502, + 0.0008203302277252078, + 0.0015145534416660666, + 0.0010844895150512457, + 0.000797649088781327, + 0.0005097538232803345, + 0.0009665600373409688, + 0.0014236539136618376, + 0.0021699487697333097, + 0.002945475047454238, + 0.0127643384039402, + 0.0035584433935582638, + 0.001726967515423894, + 0.0013784191105514765, + 0.0030869615729898214, + 0.01074429601430893, + 0.001177336904220283, + 0.0021969014778733253, + 0.0034651809837669134, + 0.0031564761884510517, + 0.0016817274736240506, + 0.0005907203303650022, + 0.0023080832324922085, + 0.002227288205176592, + 0.006119939032942057, + 0.01123822946101427, + 0.056839339435100555, + 0.005961306858807802, + 0.003726667957380414, + 0.0034824712201952934, + 0.009933155961334705, + 0.04578723758459091, + 0.0026557468809187412, + 0.0025529973208904266, + 0.0035885360557585955, + 0.0017212405800819397, + 0.0017324256477877498, + 0.0015294060576707125, + 0.0024097340647131205, + 0.004026883281767368, + 0.03588824346661568, + 0.05122501403093338, + 0.29062628746032715, + 0.02789888344705105, + 0.01765582524240017, + 0.017336737364530563, + 0.05175245180726051, + 0.21895316243171692, + 0.004124863538891077, + 0.0037920286413282156, + 0.0065954034216701984, + 0.002131393179297447, + 0.0019134684698656201, + 0.002244446659460664, + 0.002645068569108844, + 0.006471000146120787, + 0.000565249240025878, + 0.0006802970310673118, + 0.0010234757792204618, + 0.0006422544247470796, + 0.0007714012754149735, + 0.00041409238474443555, + 0.0011567323235794902, + 0.0010937926126644015, + 0.0009945154888555408, + 0.0008976901299320161, + 0.0009287532884627581, + 0.0009799666004255414, + 0.000664653955027461, + 0.0005015489878132939, + 0.0012246787082403898, + 0.0008507106686010957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.023626580834388733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255148183554411 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.01255148183554411, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255148183554411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07570015639066696, + 0.03514622524380684, + 0.6295697689056396, + 0.03591921925544739, + 0.13597406446933746, + 0.014466083608567715, + 0.030676644295454025, + 0.01529480516910553, + 0.019835609942674637, + 0.007417413871735334 + ], + "credit_map": [ + 0.0047560869716107845, + 0.00780707411468029, + 0.013107583858072758, + 0.010160690173506737, + 0.009005225263535976, + 0.003631357569247484, + 0.014544331468641758, + 0.012687805108726025, + 0.0023156076204031706, + 0.0025041992776095867, + 0.007812748663127422, + 0.004770433530211449, + 0.0036552748642861843, + 0.0016813382972031832, + 0.005572686903178692, + 0.006833935156464577, + 0.030483948066830635, + 0.017126737162470818, + 0.2641798257827759, + 0.023485975340008736, + 0.017397748306393623, + 0.01606052741408348, + 0.044183533638715744, + 0.21665148437023163, + 0.002260445384308696, + 0.0035352203994989395, + 0.009038517251610756, + 0.002691223518922925, + 0.002538225846365094, + 0.0013469572877511382, + 0.004497174639254808, + 0.010011452250182629, + 0.006083412561565638, + 0.01146737765520811, + 0.052256178110837936, + 0.0062581198289990425, + 0.0036485164891928434, + 0.0034908296074718237, + 0.00925430003553629, + 0.04351533204317093, + 0.001951511250808835, + 0.0019882917404174805, + 0.0025372267700731754, + 0.0010267402976751328, + 0.001425402588211, + 0.0009847988840192556, + 0.0015427237376570702, + 0.0030093889217823744, + 0.0020381726790219545, + 0.002538671251386404, + 0.0094003826379776, + 0.003818596014752984, + 0.002055163960903883, + 0.0010960138170048594, + 0.0035428705159574747, + 0.0061867740005254745, + 0.0019660049583762884, + 0.001697311643511057, + 0.003259469522163272, + 0.001003799494355917, + 0.001196867087855935, + 0.0011378841008991003, + 0.00150411541108042, + 0.0035293533001095057, + 0.001385235576890409, + 0.0014353275764733553, + 0.004168987739831209, + 0.002885257126763463, + 0.0018466563196852803, + 0.0010669012553989887, + 0.0029045511037111282, + 0.004142692778259516, + 0.0007625727448612452, + 0.0006052252138033509, + 0.0011151217622682452, + 0.001315007102675736, + 0.0008327395771630108, + 0.00043578012264333665, + 0.0013202936388552189, + 0.0010306740878149867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.01255148183554411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210696332156658 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210696332156658, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210696332156658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025916820392012596, + 0.011425359174609184, + 0.14343486726284027, + 0.00898278784006834, + 0.13064537942409515, + 0.018023479729890823, + 0.6305059790611267, + 0.016422299668192863, + 0.006161769852042198, + 0.00848120916634798 + ], + "credit_map": [ + 0.003007922088727355, + 0.003893553977832198, + 0.005171851720660925, + 0.002486700424924493, + 0.0019396073184907436, + 0.001874928013421595, + 0.0036304208915680647, + 0.00391183840110898, + 0.0010525177931413054, + 0.0015801662812009454, + 0.0023269273806363344, + 0.000992251094430685, + 0.0008640724117867649, + 0.0007421962800435722, + 0.0015583047643303871, + 0.002308923052623868, + 0.006820268929004669, + 0.003850203473120928, + 0.06078828498721123, + 0.005038115661591291, + 0.004013149067759514, + 0.0036977820564061403, + 0.009964540600776672, + 0.049262527376413345, + 0.0010029636323451996, + 0.0011853929609060287, + 0.0016924217343330383, + 0.0014894960913807154, + 0.0008524341974407434, + 0.0005145185859873891, + 0.0011184379691258073, + 0.001127124298363924, + 0.005693608894944191, + 0.010169748216867447, + 0.05163969472050667, + 0.005554416682571173, + 0.0033768387511372566, + 0.0031895514111965895, + 0.008948693051934242, + 0.0420728363096714, + 0.002776969922706485, + 0.0026283646002411842, + 0.0032063492108136415, + 0.001261380617506802, + 0.0015383533900603652, + 0.001407014555297792, + 0.0017921354155987501, + 0.003412913763895631, + 0.03182436525821686, + 0.0452038049697876, + 0.2575971186161041, + 0.024804797023534775, + 0.015573537908494473, + 0.015407411381602287, + 0.0458386167883873, + 0.19425639510154724, + 0.00231862161308527, + 0.0021258702035993338, + 0.002938944613561034, + 0.0013497167965397239, + 0.0014442972606047988, + 0.0013954292517155409, + 0.0016550836153328419, + 0.003194337710738182, + 0.0008067187154665589, + 0.0007825554930604994, + 0.0010011327685788274, + 0.0006264816620387137, + 0.0007746861083433032, + 0.0004746245685964823, + 0.0008933209464885294, + 0.0008022495894692838, + 0.0009185863309539855, + 0.0009563254425302148, + 0.0013045817613601685, + 0.0011367595288902521, + 0.0009009598870761693, + 0.0004872615390922874, + 0.001571703003719449, + 0.00120503269135952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.014210696332156658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798436909914017 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.029798436909914017, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798436909914017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041175905615091324, + 0.03591260313987732, + 0.1868831068277359, + 0.025414805859327316, + 0.08708735555410385, + 0.027187570929527283, + 0.5569329857826233, + 0.018449874594807625, + 0.011627646163105965, + 0.00932809617370367 + ], + "credit_map": [ + 0.003627057420089841, + 0.005085786338895559, + 0.007099898997694254, + 0.004744782578200102, + 0.004013039171695709, + 0.0020449128933250904, + 0.006668158806860447, + 0.00789226870983839, + 0.002708191517740488, + 0.003228307468816638, + 0.007205564063042402, + 0.005150668788701296, + 0.003242452396079898, + 0.0018592835403978825, + 0.004833124112337828, + 0.007685009855777025, + 0.009424013085663319, + 0.005747962277382612, + 0.07552272081375122, + 0.008113338612020016, + 0.006616945844143629, + 0.004933278542011976, + 0.014505947008728981, + 0.06201888993382454, + 0.0019766867626458406, + 0.002812150400131941, + 0.004415221977978945, + 0.0024822037667036057, + 0.002551715588197112, + 0.0011300394544377923, + 0.004013131372630596, + 0.0060336533933877945, + 0.00367232714779675, + 0.006882555782794952, + 0.03388884291052818, + 0.0039245313964784145, + 0.002631878014653921, + 0.0022357190027832985, + 0.00600938918069005, + 0.02784211002290249, + 0.003901035524904728, + 0.003340562107041478, + 0.004963954444974661, + 0.002359583042562008, + 0.0024636692833155394, + 0.002010610420256853, + 0.0029563705902546644, + 0.005191782023757696, + 0.027965359389781952, + 0.039694637060165405, + 0.22763429582118988, + 0.02201748825609684, + 0.013892270624637604, + 0.013596192933619022, + 0.04082295671105385, + 0.17130982875823975, + 0.0026424170937389135, + 0.0024525360204279423, + 0.0026511314790695906, + 0.001801034202799201, + 0.0018641912611201406, + 0.0015712613239884377, + 0.0024057182017713785, + 0.003061584196984768, + 0.001059459988027811, + 0.0011806234251707792, + 0.0021936751436442137, + 0.0013740618014708161, + 0.0011766611132770777, + 0.0006832592189311981, + 0.0019732466898858547, + 0.001986658200621605, + 0.0009426090982742608, + 0.0009619639022275805, + 0.0012365097645670176, + 0.0014543879078701138, + 0.001105933217331767, + 0.0005544005543924868, + 0.0017581814900040627, + 0.0013141086092218757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.029798436909914017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young monk with a shaved head sitting cross-legged in the lotus position in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a simple orange-red cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the monk's hands resting in front of him in a meditation mudra?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in an ancient Zen room with a dark brown wooden floor and an antique wooden lattice window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818074844777584 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.01818074844777584, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818074844777584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1257716864347458, + 0.040435194969177246, + 0.17484651505947113, + 0.024404145777225494, + 0.05307634174823761, + 0.01583360880613327, + 0.5184693336486816, + 0.016309618949890137, + 0.013718678615987301, + 0.017134934663772583 + ], + "credit_map": [ + 0.01351412944495678, + 0.019673926755785942, + 0.025380419567227364, + 0.01038734707981348, + 0.00836794264614582, + 0.010639376007020473, + 0.01286592148244381, + 0.02494259923696518, + 0.003611618187278509, + 0.0052953967824578285, + 0.008777515962719917, + 0.0040649473667144775, + 0.002542949514463544, + 0.0027250151615589857, + 0.00363534246571362, + 0.009782403707504272, + 0.008322455920279026, + 0.004845793824642897, + 0.07371264696121216, + 0.006194452755153179, + 0.0051602148450911045, + 0.004585751798003912, + 0.012000443413853645, + 0.060024719685316086, + 0.0018124459311366081, + 0.0026058251969516277, + 0.004153847694396973, + 0.00356430746614933, + 0.0026028656866401434, + 0.0013840897008776665, + 0.004287158139050007, + 0.003993602003902197, + 0.0042478665709495544, + 0.00717229163274169, + 0.014251459389925003, + 0.003386000171303749, + 0.002314555924385786, + 0.0026390363927930593, + 0.004179568029940128, + 0.014885556884109974, + 0.0020589521154761314, + 0.002816668711602688, + 0.002649510744959116, + 0.0015397059032693505, + 0.001632845145650208, + 0.0009081902680918574, + 0.001816767267882824, + 0.002410965273156762, + 0.025893500074744225, + 0.038116686046123505, + 0.21084409952163696, + 0.02045184187591076, + 0.012803807854652405, + 0.012898258864879608, + 0.03687281906604767, + 0.16058821976184845, + 0.0016281475545838475, + 0.0014395419275388122, + 0.005703473463654518, + 0.0010334240505471826, + 0.0007419848116114736, + 0.0006349284085445106, + 0.0013789405347779393, + 0.0037491759285330772, + 0.0014841649681329727, + 0.0016295283567160368, + 0.002382183214649558, + 0.001364204683341086, + 0.001023514079861343, + 0.0011574706295505166, + 0.0017272754339501262, + 0.0029503335244953632, + 0.0018752021715044975, + 0.0030089060310274363, + 0.0024058916606009007, + 0.0018679737113416195, + 0.0015271599404513836, + 0.0012213879963383079, + 0.002295380923897028, + 0.002933027222752571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.01818074844777584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_tensor([2947, 2947])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_tensor([2947, 2947])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9bdb1b73cd1384882f19a0bf3d53da728aaa1271 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_tensor([2947, 2947])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, extreme close-up meticulously captures the dramatic fluid event unfolding on a rough, gray concrete floor. A plain, unlabeled metal paint can, its silver surface slightly scuffed and slightly dented by the rim, tilts at a precarious angle. From its open mouth, a thick, viscous stream of vibrant crimson pigment cascades downward in smooth, unbroken ribbons, catching the harsh, direct overhead light and creating brilliant white highlights along its curved surface. The stream strikes the concrete with a powerful splash, and upon contact, the coalesced stream of paint shatters, erupting in a spectacular, high-speed, radial explosion. The central pool rapidly expands, ejecting countless tiny, perfectly spherical droplets in an exquisite, frozen-in-time corona. Each droplet, a microscopic sphere of pure, saturated red, contrasts against the soft, textured background, creating an intense and captivating atmosphere that isolates the raw, physical beauty of fluid dynamics in motion.", + "seed": 2985, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2985_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.26237389743328093, + "vlm_logprob": -0.26237389743328093, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.26237389743328093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08046707510948181 + ], + "opposite_logprob_tokens": [ + -3.3304669857025146 + ], + "logprob_reward": -0.08046707510948181, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08046707510948181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17895148694515228, + 0.05228402093052864, + 0.04363751411437988, + 0.059889864176511765, + 0.0730350986123085, + 0.02812257967889309, + 0.1505538523197174, + 0.27186456322669983, + 0.05109807476401329, + 0.09056287258863449 + ], + "credit_map": [ + 0.035603802651166916, + 0.039937324821949005, + 0.021965909749269485, + 0.021889209747314453, + 0.017581600695848465, + 0.012825187295675278, + 0.020051557570695877, + 0.009096893481910229, + 0.007093701511621475, + 0.008703631348907948, + 0.005850554443895817, + 0.006577640771865845, + 0.0070179360918700695, + 0.006099006161093712, + 0.005893424618989229, + 0.005048122722655535, + 0.0062119862996041775, + 0.006871085148304701, + 0.00574655132368207, + 0.004854977130889893, + 0.006391885224729776, + 0.004903830122202635, + 0.004125101491808891, + 0.004532096441835165, + 0.008989349007606506, + 0.010657641105353832, + 0.0112612908706069, + 0.009093080647289753, + 0.005475927144289017, + 0.004836225416511297, + 0.0039155976846814156, + 0.005660753231495619, + 0.019726816564798355, + 0.005919061601161957, + 0.017851989716291428, + 0.005871322937309742, + 0.006682524457573891, + 0.0055614965967834, + 0.00568592781201005, + 0.005735956132411957, + 0.0039707147516310215, + 0.0026209780480712652, + 0.003103268099948764, + 0.0022895303554832935, + 0.0032996223308146, + 0.00584837980568409, + 0.0037488010711967945, + 0.0032412863802164793, + 0.04024587571620941, + 0.04480866342782974, + 0.011608882807195187, + 0.010030308738350868, + 0.014982962980866432, + 0.009165477007627487, + 0.008060798980295658, + 0.011650893837213516, + 0.03831706568598747, + 0.11483018845319748, + 0.014213661663234234, + 0.0193426962941885, + 0.05121810361742973, + 0.015076800249516964, + 0.010962286032736301, + 0.00790377426892519, + 0.005201171152293682, + 0.013387885875999928, + 0.003345693927258253, + 0.00928039476275444, + 0.006237087305635214, + 0.005778370425105095, + 0.004994631744921207, + 0.002872840501368046, + 0.011101843789219856, + 0.013069012202322483, + 0.0065277134999632835, + 0.013667314313352108, + 0.01132144220173359, + 0.010050637647509575, + 0.017360758036375046, + 0.007464153692126274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.08046707510948181 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576251745223999 + ], + "opposite_logprob_tokens": [ + -0.826251745223999 + ], + "logprob_reward": -0.576251745223999, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576251745223999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06170928478240967, + 0.048804689198732376, + 0.02690984681248665, + 0.03309115394949913, + 0.07339296489953995, + 0.03931387886404991, + 0.02839801087975502, + 0.28505751490592957, + 0.13870573043823242, + 0.2646169066429138 + ], + "credit_map": [ + 0.008655576966702938, + 0.012168992310762405, + 0.0066459085792303085, + 0.007610409986227751, + 0.0066453274339437485, + 0.005508998408913612, + 0.008095781318843365, + 0.006378302350640297, + 0.010180807672441006, + 0.008749028667807579, + 0.005215717013925314, + 0.005741109140217304, + 0.0048819491639733315, + 0.004832681268453598, + 0.004984474275261164, + 0.00421892711892724, + 0.004687116947025061, + 0.0043113138526678085, + 0.003942617680877447, + 0.0033718040212988853, + 0.003160921623930335, + 0.0028348653577268124, + 0.0019487851532176137, + 0.002652424154803157, + 0.005069047212600708, + 0.005592401139438152, + 0.004719919990748167, + 0.003692168742418289, + 0.0054182615131139755, + 0.0034773526713252068, + 0.002295007463544607, + 0.0028269963804632425, + 0.008439239114522934, + 0.0076267602853477, + 0.007933154702186584, + 0.01150142028927803, + 0.009530831128358841, + 0.00990988314151764, + 0.011232960037887096, + 0.007218729238957167, + 0.006426996551454067, + 0.004101661033928394, + 0.005403214134275913, + 0.0024435296654701233, + 0.0048141502775251865, + 0.006104320287704468, + 0.00531132472679019, + 0.004708686377853155, + 0.003961698152124882, + 0.004282040521502495, + 0.003976630046963692, + 0.0022610926534980536, + 0.004856265150010586, + 0.003359605558216572, + 0.0023243038449436426, + 0.003376380540430546, + 0.03766648471355438, + 0.0970705971121788, + 0.016696210950613022, + 0.020227517932653427, + 0.0464276559650898, + 0.026997322216629982, + 0.018434900790452957, + 0.021536851301789284, + 0.013461541384458542, + 0.018979007378220558, + 0.012850877828896046, + 0.020956330001354218, + 0.018409134820103645, + 0.019888317212462425, + 0.02412000484764576, + 0.01004052720963955, + 0.03065023571252823, + 0.036447081714868546, + 0.020265936851501465, + 0.040183763951063156, + 0.033733949065208435, + 0.03090689890086651, + 0.05082966759800911, + 0.021599408239126205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.576251745223999 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632815420627594 + ], + "opposite_logprob_tokens": [ + -0.757815420627594 + ], + "logprob_reward": -0.632815420627594, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632815420627594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08846207708120346, + 0.050301581621170044, + 0.11778026819229126, + 0.1317921280860901, + 0.03218306973576546, + 0.05992008373141289, + 0.27911442518234253, + 0.0984986200928688, + 0.10719146579504013, + 0.03475622087717056 + ], + "credit_map": [ + 0.010003957897424698, + 0.01786811463534832, + 0.011912809684872627, + 0.011758423410356045, + 0.00923949759453535, + 0.0065169064328074455, + 0.016007505357265472, + 0.005154854152351618, + 0.006979743950068951, + 0.008250883780419827, + 0.006903527770191431, + 0.011895167641341686, + 0.004680646117776632, + 0.003782706568017602, + 0.005076540168374777, + 0.0027323667891323566, + 0.018609778955578804, + 0.025256190448999405, + 0.019679203629493713, + 0.012117546051740646, + 0.011772653087973595, + 0.011354435235261917, + 0.007929547689855099, + 0.011060911230742931, + 0.03634696081280708, + 0.020328456535935402, + 0.02090579643845558, + 0.010873819701373577, + 0.01282346248626709, + 0.009546343237161636, + 0.006249146070331335, + 0.014718150720000267, + 0.005353202577680349, + 0.0034711004700511694, + 0.004071228206157684, + 0.0024162440095096827, + 0.0028263647109270096, + 0.006963782012462616, + 0.0031485380604863167, + 0.0039326101541519165, + 0.00806465558707714, + 0.008657378144562244, + 0.009234209544956684, + 0.006015789695084095, + 0.007642919197678566, + 0.009286425076425076, + 0.006186394020915031, + 0.004832309205085039, + 0.08397532999515533, + 0.08112964034080505, + 0.020452536642551422, + 0.018247298896312714, + 0.025600286200642586, + 0.017930319532752037, + 0.0158168226480484, + 0.01596219651401043, + 0.009116877801716328, + 0.013124953024089336, + 0.01637583039700985, + 0.009692877531051636, + 0.013394973240792751, + 0.013998482376337051, + 0.011221415363252163, + 0.011573215946555138, + 0.009367920458316803, + 0.018520040437579155, + 0.010540231131017208, + 0.01615845412015915, + 0.014144981279969215, + 0.016532085835933685, + 0.014669310301542282, + 0.007258444093167782, + 0.003993694204837084, + 0.006254762411117554, + 0.003438112558797002, + 0.0050979130901396275, + 0.003493795869871974, + 0.00398018304258585, + 0.0053826672956347466, + 0.0031150926370173693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.632815420627594 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553011998534203 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553011998534203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553011998534203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08406516164541245, + 0.04470257833600044, + 0.1337297409772873, + 0.04919978603720665, + 0.05709460750222206, + 0.0434231236577034, + 0.09346267580986023, + 0.11692921072244644, + 0.10873143374919891, + 0.2686617076396942 + ], + "credit_map": [ + 0.010744784027338028, + 0.014794747345149517, + 0.010848010890185833, + 0.011222920380532742, + 0.008878747932612896, + 0.006895462516695261, + 0.01582765206694603, + 0.004852833226323128, + 0.005926727317273617, + 0.0069291251711547375, + 0.004875323735177517, + 0.007251888047903776, + 0.006498867645859718, + 0.004639686085283756, + 0.004936649464070797, + 0.003644306445494294, + 0.02196940779685974, + 0.030315931886434555, + 0.024578221142292023, + 0.016470806673169136, + 0.013037092983722687, + 0.008512292988598347, + 0.010507402010262012, + 0.0083385668694973, + 0.008694227784872055, + 0.011111707426607609, + 0.007752397563308477, + 0.006804886274039745, + 0.0037514064460992813, + 0.003359587863087654, + 0.004161412827670574, + 0.0035641551949083805, + 0.016412418335676193, + 0.004668214358389378, + 0.013964260928332806, + 0.004688830580562353, + 0.00424594059586525, + 0.004508343990892172, + 0.003848560154438019, + 0.004758037626743317, + 0.00752983195707202, + 0.005149426870048046, + 0.007028115447610617, + 0.00313957198522985, + 0.005069839768111706, + 0.006906447932124138, + 0.004268252290785313, + 0.004331633448600769, + 0.019803404808044434, + 0.02575674094259739, + 0.008847557008266449, + 0.0067085446789860725, + 0.00958006177097559, + 0.006306851748377085, + 0.005370895843952894, + 0.011088603176176548, + 0.009173469617962837, + 0.015773991122841835, + 0.018480123952031136, + 0.012106867507100105, + 0.012868610210716724, + 0.018101515248417854, + 0.015089629217982292, + 0.015334982424974442, + 0.01174958050251007, + 0.0293520949780941, + 0.006765836849808693, + 0.022853003814816475, + 0.00985952652990818, + 0.01054627075791359, + 0.011984550394117832, + 0.00562055641785264, + 0.030587032437324524, + 0.033351629972457886, + 0.020786788314580917, + 0.041120439767837524, + 0.034544799476861954, + 0.032872654497623444, + 0.05215408653020859, + 0.023244228214025497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.012553011998534203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782234206795692 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782234206795692, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782234206795692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702641397714615, + 0.14760270714759827, + 0.11312127113342285, + 0.11939150094985962, + 0.12492412328720093, + 0.05662359297275543, + 0.06814278662204742, + 0.05849156528711319, + 0.0774148479104042, + 0.0640234500169754 + ], + "credit_map": [ + 0.022848671302199364, + 0.028849447146058083, + 0.024921441450715065, + 0.02363584004342556, + 0.01643422432243824, + 0.01292932778596878, + 0.03089475817978382, + 0.009750442579388618, + 0.021257145330309868, + 0.024697260931134224, + 0.015811825171113014, + 0.023300355300307274, + 0.018395915627479553, + 0.01702544279396534, + 0.014080130495131016, + 0.013034630566835403, + 0.023448383435606956, + 0.022091401740908623, + 0.018470395356416702, + 0.011278567835688591, + 0.0112941674888134, + 0.010346692986786366, + 0.006638199090957642, + 0.009553461335599422, + 0.03075387515127659, + 0.018942460417747498, + 0.017424944788217545, + 0.011321957223117352, + 0.01338668167591095, + 0.010227273218333721, + 0.006930953357368708, + 0.010403355583548546, + 0.04602989926934242, + 0.0073319897055625916, + 0.04138561710715294, + 0.006483240053057671, + 0.0058511109091341496, + 0.007947782054543495, + 0.004753894172608852, + 0.005140599329024553, + 0.012650121003389359, + 0.006133603863418102, + 0.011758888140320778, + 0.003328960156068206, + 0.005629871040582657, + 0.007527013309299946, + 0.005376625340431929, + 0.0042185103520751, + 0.016478247940540314, + 0.019737888127565384, + 0.006329953204840422, + 0.004158031195402145, + 0.00686922762542963, + 0.004188538528978825, + 0.003246195148676634, + 0.007134702056646347, + 0.004900411237031221, + 0.017208561301231384, + 0.006433524191379547, + 0.00432620057836175, + 0.008708702400326729, + 0.0070870728231966496, + 0.005039141047745943, + 0.004787949379533529, + 0.008962028659880161, + 0.013504012487828732, + 0.0066947634331882, + 0.015668194741010666, + 0.007550724782049656, + 0.010709925554692745, + 0.009328072890639305, + 0.00499712722375989, + 0.007026844657957554, + 0.0059701232239604, + 0.005741807632148266, + 0.01004855241626501, + 0.007500643841922283, + 0.009563510306179523, + 0.01214243657886982, + 0.0060295299626886845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009782234206795692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up meticulously captures the dramatic fluid event unfolding on a rough, gray concrete floor. A plain, unlabeled metal paint can, its silver surface slightly scuffed and slightly dented by the rim, tilts at a precarious angle. From its open mouth, a thick, viscous stream of vibrant crimson pigment cascades downward in smooth, unbroken ribbons, catching the harsh, direct overhead light and creating brilliant white highlights along its curved surface. The stream strikes the concrete with a powerful splash, and upon contact, the coalesced stream of paint shatters, erupting in a spectacular, high-speed, radial explosion. The central pool rapidly expands, ejecting countless tiny, perfectly spherical droplets in an exquisite, frozen-in-time corona. Each droplet, a microscopic sphere of pure, saturated red, contrasts against the soft, textured background, creating an intense and captivating atmosphere that isolates the raw, physical beauty of fluid dynamics in motion.", + "seed": 2986, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2986_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0748916900716722, + "vlm_logprob": -0.0748916900716722, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.0748916900716722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23568832874298096 + ], + "opposite_logprob_tokens": [ + -1.735688328742981 + ], + "logprob_reward": -0.23568832874298096, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23568832874298096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39959657192230225, + 0.042610667645931244, + 0.050443630665540695, + 0.04067324474453926, + 0.06542132794857025, + 0.05625767260789871, + 0.08637962490320206, + 0.1580480933189392, + 0.03294000029563904, + 0.06762918829917908 + ], + "credit_map": [ + 0.05794144794344902, + 0.05125755816698074, + 0.07639344781637192, + 0.04260313883423805, + 0.04952951520681381, + 0.03908326104283333, + 0.04290086403489113, + 0.03988729789853096, + 0.006729886867105961, + 0.006617972627282143, + 0.006961369421333075, + 0.004376290366053581, + 0.003276367438957095, + 0.005026827100664377, + 0.006804326083511114, + 0.0028176249470561743, + 0.006120174191892147, + 0.009024804458022118, + 0.006135343108326197, + 0.007727427873760462, + 0.00416136160492897, + 0.006304990500211716, + 0.006768880877643824, + 0.0042006466537714005, + 0.005155737511813641, + 0.0064489571377635, + 0.005725855007767677, + 0.005263868253678083, + 0.004726724699139595, + 0.006536547094583511, + 0.003631629515439272, + 0.0031839250586926937, + 0.010043352842330933, + 0.011126858182251453, + 0.009555475786328316, + 0.009185219183564186, + 0.004182584583759308, + 0.009862160310149193, + 0.006162903271615505, + 0.005302773788571358, + 0.009642226621508598, + 0.008102170191705227, + 0.006837885361164808, + 0.006658731494098902, + 0.005011287983506918, + 0.006305687595158815, + 0.00810528639703989, + 0.005594391375780106, + 0.01104514766484499, + 0.01228822860866785, + 0.018664216622710228, + 0.007940253242850304, + 0.009143706411123276, + 0.0100402245298028, + 0.01019144244492054, + 0.007066401652991772, + 0.015426037833094597, + 0.02008829079568386, + 0.010253801010549068, + 0.009636485949158669, + 0.048533931374549866, + 0.02455008774995804, + 0.018370984122157097, + 0.011188461445271969, + 0.0032552951015532017, + 0.006656270008534193, + 0.003467448288574815, + 0.0023216325789690018, + 0.005081261973828077, + 0.0044668312184512615, + 0.003845680970698595, + 0.003845576662570238, + 0.006456158589571714, + 0.008350121788680553, + 0.005453122779726982, + 0.006539769936352968, + 0.01578095369040966, + 0.00683177076280117, + 0.011538976803421974, + 0.006678315810859203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.23568832874298096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035981237888336 + ], + "opposite_logprob_tokens": [ + -2.3503599166870117 + ], + "logprob_reward": -0.10035981237888336, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035981237888336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15252842009067535, + 0.05152066424489021, + 0.052724722772836685, + 0.0727703720331192, + 0.06430072337388992, + 0.046489179134368896, + 0.09569207578897476, + 0.34664174914360046, + 0.025049515068531036, + 0.09228263795375824 + ], + "credit_map": [ + 0.025656208395957947, + 0.01664409041404724, + 0.03885399177670479, + 0.021839331835508347, + 0.0089360186830163, + 0.013850545510649681, + 0.013803509995341301, + 0.012944718822836876, + 0.00772266136482358, + 0.005019558593630791, + 0.008877143263816833, + 0.0051854695193469524, + 0.0028243130072951317, + 0.011032728478312492, + 0.006750428583472967, + 0.0041083600372076035, + 0.009271087124943733, + 0.008634839206933975, + 0.007036332506686449, + 0.007429947145283222, + 0.00312354089692235, + 0.007071297615766525, + 0.004753125365823507, + 0.005404547788202763, + 0.009055170230567455, + 0.009368774481117725, + 0.011869038455188274, + 0.009077678434550762, + 0.007282246369868517, + 0.01099887304008007, + 0.007568407338112593, + 0.007550181820988655, + 0.011073286645114422, + 0.010605218820273876, + 0.0088818883523345, + 0.006329711060971022, + 0.006115153897553682, + 0.012404491193592548, + 0.0042646499350667, + 0.004626323003321886, + 0.009792168624699116, + 0.007941617630422115, + 0.006670638453215361, + 0.0031488456297665834, + 0.005499956663697958, + 0.005059160757809877, + 0.004651347175240517, + 0.0037254439666867256, + 0.012282238341867924, + 0.010089637711644173, + 0.01750430278480053, + 0.009840868413448334, + 0.011812549084424973, + 0.01576501689851284, + 0.011286333203315735, + 0.007111126556992531, + 0.03335994854569435, + 0.03700102120637894, + 0.024473266676068306, + 0.01967777870595455, + 0.11365888267755508, + 0.05611970275640488, + 0.038711413741111755, + 0.023639723658561707, + 0.002387631917372346, + 0.004046538379043341, + 0.0018784370040521026, + 0.0022405912168323994, + 0.00445556128397584, + 0.002752084983512759, + 0.003302254481241107, + 0.003986415918916464, + 0.008783867582678795, + 0.009871632792055607, + 0.005873630754649639, + 0.006889664102345705, + 0.023304250091314316, + 0.011409761384129524, + 0.015447982586920261, + 0.010701848194003105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.10035981237888336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822569966316223 + ], + "opposite_logprob_tokens": [ + -4.01822566986084 + ], + "logprob_reward": -0.01822569966316223, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822569966316223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2900201082229614, + 0.06330999732017517, + 0.1043999046087265, + 0.047591786831617355, + 0.1194043904542923, + 0.05641709268093109, + 0.1071622222661972, + 0.10701385885477066, + 0.03815549239516258, + 0.06652508676052094 + ], + "credit_map": [ + 0.04582221060991287, + 0.03800029680132866, + 0.06117922440171242, + 0.0364919938147068, + 0.03079257346689701, + 0.025560008361935616, + 0.026298098266124725, + 0.025875728577375412, + 0.009446486830711365, + 0.013703614473342896, + 0.01085809525102377, + 0.006190838757902384, + 0.0043216426856815815, + 0.007997225970029831, + 0.006843335926532745, + 0.0039487555623054504, + 0.013413106091320515, + 0.014329446479678154, + 0.0137476259842515, + 0.011210592463612556, + 0.009888614527881145, + 0.020462987944483757, + 0.009768611751496792, + 0.011578925885260105, + 0.0065581039525568485, + 0.006566266994923353, + 0.006178035866469145, + 0.006369455251842737, + 0.004270642530173063, + 0.008057737722992897, + 0.005975867621600628, + 0.003615677822381258, + 0.014264681376516819, + 0.011939515359699726, + 0.014442597515881062, + 0.017530713230371475, + 0.019895758479833603, + 0.01776149682700634, + 0.013968340121209621, + 0.009601294994354248, + 0.011256602592766285, + 0.009005975909531116, + 0.007572667673230171, + 0.0064783645793795586, + 0.005236120894551277, + 0.006987303961068392, + 0.005079459864646196, + 0.004800597205758095, + 0.01457631029188633, + 0.010478953830897808, + 0.01722170226275921, + 0.014326600357890129, + 0.008794580586254597, + 0.011911765672266483, + 0.02016923762857914, + 0.009683080948889256, + 0.010266509838402271, + 0.014292126521468163, + 0.008264578878879547, + 0.006058094557374716, + 0.0328451506793499, + 0.017025144770741463, + 0.011814036406576633, + 0.006448219995945692, + 0.003725286340340972, + 0.005260877311229706, + 0.004430756438523531, + 0.003115080064162612, + 0.00678006699308753, + 0.005174203310161829, + 0.004572628997266293, + 0.005096597131341696, + 0.006292041856795549, + 0.009021222591400146, + 0.004174578003585339, + 0.004882067441940308, + 0.013691170141100883, + 0.009554190561175346, + 0.011935503222048283, + 0.006974322255700827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.01822569966316223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640897762030363 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007640897762030363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640897762030363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11078961938619614, + 0.045569367706775665, + 0.06933894008398056, + 0.08583152294158936, + 0.06144974008202553, + 0.0838443711400032, + 0.07080112397670746, + 0.3596291244029999, + 0.04011692851781845, + 0.07262929528951645 + ], + "credit_map": [ + 0.01931711658835411, + 0.014399705454707146, + 0.027126824483275414, + 0.01413688063621521, + 0.00899828877300024, + 0.009618937969207764, + 0.008978519588708878, + 0.008213341236114502, + 0.006781199015676975, + 0.0060001336969435215, + 0.007041655480861664, + 0.005797307938337326, + 0.002762605668976903, + 0.00688981031998992, + 0.006828033830970526, + 0.00346861919388175, + 0.009924850426614285, + 0.010408660396933556, + 0.008535053580999374, + 0.01341693289577961, + 0.0035776467993855476, + 0.007721634116023779, + 0.009504657238721848, + 0.006249506026506424, + 0.013077410869300365, + 0.011856461875140667, + 0.017527883872389793, + 0.007833577692508698, + 0.007150772027671337, + 0.012868239544332027, + 0.007389862556010485, + 0.008127318695187569, + 0.006790607701987028, + 0.005377587396651506, + 0.010231695137917995, + 0.006448633037507534, + 0.010677415877580643, + 0.011034267023205757, + 0.006168065592646599, + 0.004721467848867178, + 0.012691905722022057, + 0.009571562521159649, + 0.009798740036785603, + 0.01314463373273611, + 0.009817700833082199, + 0.0108340410515666, + 0.010250254534184933, + 0.0077355303801596165, + 0.010010636411607265, + 0.010570360347628593, + 0.012339508160948753, + 0.006406232248991728, + 0.00851772353053093, + 0.007040097378194332, + 0.009252343326807022, + 0.006664218381047249, + 0.035041265189647675, + 0.03814221918582916, + 0.02674773707985878, + 0.02122841589152813, + 0.11297336965799332, + 0.058657027781009674, + 0.04230276867747307, + 0.024536313489079475, + 0.0043482230976223946, + 0.006749201565980911, + 0.005237454082816839, + 0.004903330467641354, + 0.005133495200425386, + 0.0036210177931934595, + 0.004733043722808361, + 0.005391161888837814, + 0.006968060974031687, + 0.009136748500168324, + 0.004632499534636736, + 0.00569012388586998, + 0.016729682683944702, + 0.00989469513297081, + 0.011381569318473339, + 0.008195922710001469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.007640897762030363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543711811304092 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543711811304092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543711811304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21370473504066467, + 0.04777469113469124, + 0.07398339360952377, + 0.0677105188369751, + 0.055378880351781845, + 0.07452455163002014, + 0.10067795217037201, + 0.19696691632270813, + 0.08823303878307343, + 0.0810452476143837 + ], + "credit_map": [ + 0.03718479722738266, + 0.026805037632584572, + 0.037818338721990585, + 0.023470263928174973, + 0.018366605043411255, + 0.03478424996137619, + 0.019440526142716408, + 0.015834923833608627, + 0.007293334696441889, + 0.01136974897235632, + 0.006685859058052301, + 0.005059366580098867, + 0.002798820612952113, + 0.006064346060156822, + 0.00499340845271945, + 0.003509811358526349, + 0.009451132267713547, + 0.007881264202296734, + 0.012664301320910454, + 0.011558698490262032, + 0.005189685616642237, + 0.008289436809718609, + 0.013406133279204369, + 0.005542753264307976, + 0.01101989857852459, + 0.009898378513753414, + 0.013042223639786243, + 0.0059528411366045475, + 0.006005196366459131, + 0.009025678038597107, + 0.006075295154005289, + 0.0066910176537930965, + 0.006892603822052479, + 0.007538049016147852, + 0.007731341291218996, + 0.007885510101914406, + 0.005938471760600805, + 0.008785027079284191, + 0.006019267253577709, + 0.004588617943227291, + 0.015906238928437233, + 0.011923346668481827, + 0.010939134284853935, + 0.005903102457523346, + 0.0061408220790326595, + 0.011600756086409092, + 0.006154713686555624, + 0.005956439767032862, + 0.013038389384746552, + 0.009995872154831886, + 0.017188671976327896, + 0.013128869235515594, + 0.01083717867732048, + 0.012293696403503418, + 0.015423722565174103, + 0.008771559223532677, + 0.01934739015996456, + 0.0241763424128294, + 0.015449488535523415, + 0.012920106761157513, + 0.06006001681089401, + 0.03304583951830864, + 0.01766732707619667, + 0.014300411567091942, + 0.010896904394030571, + 0.022678256034851074, + 0.013466916978359222, + 0.005895376205444336, + 0.010368308052420616, + 0.012102564796805382, + 0.0062063452787697315, + 0.006618380546569824, + 0.007932526990771294, + 0.010974487289786339, + 0.005351882893592119, + 0.0065161315724253654, + 0.018431300297379494, + 0.012461616657674313, + 0.01167110912501812, + 0.0077061960473656654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.012543711811304092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up meticulously captures the dramatic fluid event unfolding on a rough, gray concrete floor. A plain, unlabeled metal paint can, its silver surface slightly scuffed and slightly dented by the rim, tilts at a precarious angle. From its open mouth, a thick, viscous stream of vibrant crimson pigment cascades downward in smooth, unbroken ribbons, catching the harsh, direct overhead light and creating brilliant white highlights along its curved surface. The stream strikes the concrete with a powerful splash, and upon contact, the coalesced stream of paint shatters, erupting in a spectacular, high-speed, radial explosion. The central pool rapidly expands, ejecting countless tiny, perfectly spherical droplets in an exquisite, frozen-in-time corona. Each droplet, a microscopic sphere of pure, saturated red, contrasts against the soft, textured background, creating an intense and captivating atmosphere that isolates the raw, physical beauty of fluid dynamics in motion.", + "seed": 2987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2987_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.6862381726503373, + "vlm_logprob": -0.6862381726503373, + "vlm_yes_no_margin": 1.275, + "vlm_token_credit": -0.6862381726503373, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28233930468559265 + ], + "opposite_logprob_tokens": [ + -1.532339334487915 + ], + "logprob_reward": -0.28233930468559265, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.28233930468559265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10056062042713165, + 0.0986088365316391, + 0.14440229535102844, + 0.16769233345985413, + 0.17666639387607574, + 0.07335986942052841, + 0.045249093323946, + 0.05763214826583862, + 0.04528830572962761, + 0.0905400961637497 + ], + "credit_map": [ + 0.02111995965242386, + 0.015451294369995594, + 0.014233422465622425, + 0.01594194583594799, + 0.006295597646385431, + 0.008780881762504578, + 0.008087706752121449, + 0.01064981147646904, + 0.026569904759526253, + 0.013215091079473495, + 0.016116684302687645, + 0.009771371260285378, + 0.006482874974608421, + 0.011523976922035217, + 0.007022232748568058, + 0.007906703278422356, + 0.028329089283943176, + 0.016229160130023956, + 0.017712537199258804, + 0.02493496797978878, + 0.008531460538506508, + 0.02522941119968891, + 0.012285241857171059, + 0.011150429025292397, + 0.025853393599390984, + 0.01844753883779049, + 0.02631349489092827, + 0.019845904782414436, + 0.010629089549183846, + 0.037318792194128036, + 0.013629005290567875, + 0.015655117109417915, + 0.027533166110515594, + 0.020966799929738045, + 0.022424736991524696, + 0.018978051841259003, + 0.010626981034874916, + 0.04127954691648483, + 0.014042296446859837, + 0.020814813673496246, + 0.014969810843467712, + 0.005347400903701782, + 0.0066634006798267365, + 0.0066063120029866695, + 0.005096966400742531, + 0.007368520833551884, + 0.009273034520447254, + 0.018034420907497406, + 0.005367293022572994, + 0.0038336613215506077, + 0.006546481512486935, + 0.004741495009511709, + 0.007722344249486923, + 0.006761312484741211, + 0.004824343603104353, + 0.005452165845781565, + 0.006058130878955126, + 0.005450232420116663, + 0.007491082884371281, + 0.009459363296627998, + 0.005454213358461857, + 0.0081608472391963, + 0.009665807709097862, + 0.00589247141033411, + 0.004263020120561123, + 0.0037292540073394775, + 0.006298220716416836, + 0.012334388680756092, + 0.003569146618247032, + 0.00623320322483778, + 0.005378958769142628, + 0.003482114290818572, + 0.006045295391231775, + 0.005639232229441404, + 0.03425559028983116, + 0.024200908839702606, + 0.004102620296180248, + 0.0037152564618736506, + 0.007091860752552748, + 0.005489340517669916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.28233930468559265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328577399253845 + ], + "opposite_logprob_tokens": [ + -0.7578577399253845 + ], + "logprob_reward": -0.6328577399253845, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328577399253845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05667613819241524, + 0.20059199631214142, + 0.12153889238834381, + 0.04675149917602539, + 0.27673909068107605, + 0.06969814002513885, + 0.022016819566488266, + 0.04542773962020874, + 0.049072057008743286, + 0.11148759722709656 + ], + "credit_map": [ + 0.013412773609161377, + 0.008567890152335167, + 0.008205452933907509, + 0.006243641953915358, + 0.004200374241918325, + 0.0063506728038191795, + 0.004389849491417408, + 0.005305479746311903, + 0.06614162772893906, + 0.03264545649290085, + 0.03545280173420906, + 0.008377165533602238, + 0.009024026803672314, + 0.032743994146585464, + 0.005929692182689905, + 0.010277239605784416, + 0.02092853933572769, + 0.012809068895876408, + 0.012807494029402733, + 0.02336796373128891, + 0.007550118491053581, + 0.02303728088736534, + 0.011036333627998829, + 0.010002097114920616, + 0.0072682201862335205, + 0.006440026219934225, + 0.0049641188234090805, + 0.008448487147688866, + 0.003708158852532506, + 0.00827518105506897, + 0.00446498766541481, + 0.0031823187600821257, + 0.04069042578339577, + 0.028772754594683647, + 0.037091948091983795, + 0.02896149642765522, + 0.01663125678896904, + 0.06759389489889145, + 0.020898228511214256, + 0.036099113523960114, + 0.013884845189750195, + 0.004589863121509552, + 0.00767825311049819, + 0.00531220156699419, + 0.0034964894875884056, + 0.013414612039923668, + 0.005914479028433561, + 0.015407395549118519, + 0.002790795173496008, + 0.0018162417691200972, + 0.002973568858578801, + 0.00245014950633049, + 0.0029008539859205484, + 0.0034159536007791758, + 0.0029758173041045666, + 0.0026934377383440733, + 0.005028590559959412, + 0.002875392558053136, + 0.00962536409497261, + 0.010369443334639072, + 0.003633856540545821, + 0.004139469005167484, + 0.007132682949304581, + 0.002622941043227911, + 0.0037897881120443344, + 0.004136431962251663, + 0.005390964914113283, + 0.008987748064100742, + 0.00557634187862277, + 0.012247782200574875, + 0.00470096617937088, + 0.004242032766342163, + 0.006973984185606241, + 0.017757905647158623, + 0.04336833208799362, + 0.02315821498632431, + 0.0026160296984016895, + 0.00612600427120924, + 0.008601871319115162, + 0.0028852547984570265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.6328577399253845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4640274047851562 + ], + "opposite_logprob_tokens": [ + -0.08902736008167267 + ], + "logprob_reward": -2.4640274047851562, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4640274047851562, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06279683113098145, + 0.08724191039800644, + 0.07582369446754456, + 0.14866158366203308, + 0.07407160848379135, + 0.073442243039608, + 0.03973199799656868, + 0.023030349984765053, + 0.09829273074865341, + 0.3169071078300476 + ], + "credit_map": [ + 0.010140069760382175, + 0.010955397970974445, + 0.008592105470597744, + 0.007781719323247671, + 0.005351571831852198, + 0.007153850514441729, + 0.0062668826431035995, + 0.006555229425430298, + 0.023685351014137268, + 0.01309165544807911, + 0.013573662377893925, + 0.008184103295207024, + 0.004901580978184938, + 0.010396618396043777, + 0.007188152987509966, + 0.006220785900950432, + 0.01457716803997755, + 0.008637813851237297, + 0.009014604613184929, + 0.014885061420500278, + 0.004404506646096706, + 0.010243173688650131, + 0.006499037612229586, + 0.007562334183603525, + 0.02037934772670269, + 0.013902933336794376, + 0.022197192534804344, + 0.0359693206846714, + 0.007611424662172794, + 0.02290351688861847, + 0.015784507617354393, + 0.009913346730172634, + 0.011284422129392624, + 0.015315964818000793, + 0.006323545705527067, + 0.005616201553493738, + 0.004708265420049429, + 0.018340090289711952, + 0.006369910668581724, + 0.00611321022734046, + 0.01747988350689411, + 0.005707672331482172, + 0.010738290846347809, + 0.007649337872862816, + 0.0070649404078722, + 0.008584640920162201, + 0.009412992745637894, + 0.0068044839426875114, + 0.004340180195868015, + 0.0034023213665932417, + 0.008353710174560547, + 0.004539545625448227, + 0.006845229305326939, + 0.004659241531044245, + 0.004200786352157593, + 0.0033909818157553673, + 0.003025044919922948, + 0.001981759909540415, + 0.0028515432495623827, + 0.004084395244717598, + 0.0020939945243299007, + 0.0031056590378284454, + 0.0040229205042123795, + 0.0018650311976671219, + 0.00844181701540947, + 0.01173302624374628, + 0.0061660571955144405, + 0.013940863311290741, + 0.011409041471779346, + 0.03144444152712822, + 0.006943882443010807, + 0.008213602937757969, + 0.020684242248535156, + 0.0416388139128685, + 0.13120964169502258, + 0.07572174072265625, + 0.006221634801477194, + 0.012231331318616867, + 0.02277172915637493, + 0.006427940912544727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -2.4640274047851562 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01823880895972252 + ], + "opposite_logprob_tokens": [ + -4.0182390213012695 + ], + "logprob_reward": -0.01823880895972252, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01823880895972252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05737590044736862, + 0.04392346739768982, + 0.10344045609235764, + 0.1846400797367096, + 0.20396876335144043, + 0.11927700787782669, + 0.017945975065231323, + 0.04953121766448021, + 0.07366251200437546, + 0.14623461663722992 + ], + "credit_map": [ + 0.013114865869283676, + 0.009867234155535698, + 0.008427801541984081, + 0.004494897089898586, + 0.004639055114239454, + 0.006895884405821562, + 0.004997799172997475, + 0.004938362631946802, + 0.00997986365109682, + 0.00682002492249012, + 0.006800840143114328, + 0.0034162290394306183, + 0.003372445935383439, + 0.006081396713852882, + 0.0039411126635968685, + 0.0035115559585392475, + 0.02179456315934658, + 0.013150974176824093, + 0.013239150866866112, + 0.011690405197441578, + 0.00592926237732172, + 0.02530214749276638, + 0.005383332725614309, + 0.00695062568411231, + 0.027952618896961212, + 0.01929805800318718, + 0.025527918711304665, + 0.04752233624458313, + 0.010497793555259705, + 0.02392696402966976, + 0.01793006621301174, + 0.011984329670667648, + 0.03329547867178917, + 0.04154035821557045, + 0.0235088262706995, + 0.01711459830403328, + 0.014706073328852654, + 0.039936866611242294, + 0.015602968633174896, + 0.018263595178723335, + 0.027754271402955055, + 0.0106199961155653, + 0.01893991231918335, + 0.013067148625850677, + 0.00695828627794981, + 0.01325202826410532, + 0.010884967632591724, + 0.017800403758883476, + 0.00218170671723783, + 0.0017700593452900648, + 0.0029368565883487463, + 0.0018903544405475259, + 0.002719020703807473, + 0.002593090059235692, + 0.001922985422424972, + 0.0019319026032462716, + 0.006138218101114035, + 0.003428921801969409, + 0.009563066065311432, + 0.0105388592928648, + 0.004622685723006725, + 0.005222347564995289, + 0.007336587645113468, + 0.0026805317029356956, + 0.006614836864173412, + 0.009036210365593433, + 0.006382985506206751, + 0.007109587546437979, + 0.010311885736882687, + 0.022639356553554535, + 0.0055791884660720825, + 0.005988456308841705, + 0.010328290984034538, + 0.007827541790902615, + 0.06322828680276871, + 0.04213171824812889, + 0.002868880284950137, + 0.005172327626496553, + 0.010575291700661182, + 0.00410228269174695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.01823880895972252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033727604895830154 + ], + "opposite_logprob_tokens": [ + -3.4087276458740234 + ], + "logprob_reward": -0.033727604895830154, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033727604895830154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07616008073091507, + 0.10583657771348953, + 0.2436135709285736, + 0.09186778217554092, + 0.10543261468410492, + 0.1206931471824646, + 0.023786399513483047, + 0.02826228179037571, + 0.05610194802284241, + 0.14824560284614563 + ], + "credit_map": [ + 0.01838228665292263, + 0.01238766685128212, + 0.01060066930949688, + 0.008980631828308105, + 0.005510870832949877, + 0.008812188170850277, + 0.005385871510952711, + 0.00609989371150732, + 0.03167812153697014, + 0.01563199982047081, + 0.017501771450042725, + 0.006732826586812735, + 0.005589483771473169, + 0.02008766122162342, + 0.003433534409850836, + 0.005181180313229561, + 0.045550283044576645, + 0.026633245870471, + 0.02826312743127346, + 0.05226268991827965, + 0.014457972720265388, + 0.029966633766889572, + 0.024485137313604355, + 0.02199447527527809, + 0.010227654129266739, + 0.007390059996396303, + 0.012477072887122631, + 0.02871614694595337, + 0.005951560102403164, + 0.011565443128347397, + 0.010189762338995934, + 0.005350084044039249, + 0.016641173511743546, + 0.018004218116402626, + 0.010460175573825836, + 0.009567586705088615, + 0.0069669936783611774, + 0.023667771369218826, + 0.00911677721887827, + 0.01100791897624731, + 0.024047138169407845, + 0.011138338595628738, + 0.020705418661236763, + 0.01658342033624649, + 0.008368567563593388, + 0.017721358686685562, + 0.013023599050939083, + 0.009105301462113857, + 0.002664917381480336, + 0.0021298048086464405, + 0.004554548766463995, + 0.002639015205204487, + 0.003929883241653442, + 0.002901889383792877, + 0.002434410387650132, + 0.0025319296400994062, + 0.0030414725188165903, + 0.0020998490508645773, + 0.00545811140909791, + 0.006350622046738863, + 0.0025037862360477448, + 0.0027794039342552423, + 0.004362213425338268, + 0.0016668230528011918, + 0.005228913854807615, + 0.006555362604558468, + 0.003432422410696745, + 0.007604349870234728, + 0.007393117994070053, + 0.01721307635307312, + 0.003986838273704052, + 0.004687868058681488, + 0.010869679041206837, + 0.007412554696202278, + 0.06393219530582428, + 0.04151684045791626, + 0.0040859743021428585, + 0.005486041773110628, + 0.010794118978083134, + 0.004148200154304504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.033727604895830154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, extreme close-up meticulously captures the dramatic fluid event unfolding on a rough, gray concrete floor. A plain, unlabeled metal paint can, its silver surface slightly scuffed and slightly dented by the rim, tilts at a precarious angle. From its open mouth, a thick, viscous stream of vibrant crimson pigment cascades downward in smooth, unbroken ribbons, catching the harsh, direct overhead light and creating brilliant white highlights along its curved surface. The stream strikes the concrete with a powerful splash, and upon contact, the coalesced stream of paint shatters, erupting in a spectacular, high-speed, radial explosion. The central pool rapidly expands, ejecting countless tiny, perfectly spherical droplets in an exquisite, frozen-in-time corona. Each droplet, a microscopic sphere of pure, saturated red, contrasts against the soft, textured background, creating an intense and captivating atmosphere that isolates the raw, physical beauty of fluid dynamics in motion.", + "seed": 2988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_7/index_2947_seed_2988_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4040079992264509, + "vlm_logprob": -0.4040079992264509, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.4040079992264509, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25612741708755493 + ], + "opposite_logprob_tokens": [ + -1.6311273574829102 + ], + "logprob_reward": -0.25612741708755493, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.25612741708755493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2717612087726593, + 0.14377044141292572, + 0.0664592981338501, + 0.05164901167154312, + 0.03778783604502678, + 0.15551656484603882, + 0.05693351477384567, + 0.07957252115011215, + 0.08233223855495453, + 0.054217394441366196 + ], + "credit_map": [ + 0.023578383028507233, + 0.03217921033501625, + 0.032104503363370895, + 0.05287342146039009, + 0.023390986025333405, + 0.015267600305378437, + 0.07382127642631531, + 0.01854582317173481, + 0.022073030471801758, + 0.023426637053489685, + 0.02271476574242115, + 0.02501724101603031, + 0.013129952363669872, + 0.013696289621293545, + 0.013416223227977753, + 0.010296300984919071, + 0.008791766129434109, + 0.00813845545053482, + 0.008318336680531502, + 0.009124587289988995, + 0.007569531910121441, + 0.006795194931328297, + 0.009114391170442104, + 0.008607026189565659, + 0.006219420116394758, + 0.00574088841676712, + 0.005963935982435942, + 0.00942605547606945, + 0.006104451138526201, + 0.005024205893278122, + 0.005347551312297583, + 0.007822503335773945, + 0.005577538162469864, + 0.004354497417807579, + 0.0052339909598231316, + 0.0048177000135183334, + 0.0043516457080841064, + 0.005976058542728424, + 0.004066969733685255, + 0.0034094343427568674, + 0.023439163342118263, + 0.015487534925341606, + 0.020537175238132477, + 0.01084280014038086, + 0.018897026777267456, + 0.04454346001148224, + 0.00960643868893385, + 0.012162969447672367, + 0.0037712291814386845, + 0.01733797788619995, + 0.004535438027232885, + 0.004872071091085672, + 0.004610194358974695, + 0.013379955664277077, + 0.003918859641999006, + 0.004507793113589287, + 0.01115178782492876, + 0.014366858638823032, + 0.010276900604367256, + 0.009918686002492905, + 0.00907206255942583, + 0.007796861231327057, + 0.011825000867247581, + 0.005164364352822304, + 0.005262542981654406, + 0.028102301061153412, + 0.022434212267398834, + 0.00567267881706357, + 0.006540598813444376, + 0.005727802403271198, + 0.005132708698511124, + 0.0034593958407640457, + 0.008488682098686695, + 0.006315360777080059, + 0.005049569066613913, + 0.005232266150414944, + 0.008851251564919949, + 0.009681075811386108, + 0.005214151460677385, + 0.005385035183280706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.25612741708755493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761873126029968 + ], + "opposite_logprob_tokens": [ + -0.8261873126029968 + ], + "logprob_reward": -0.5761873126029968, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761873126029968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1624305546283722, + 0.1265261024236679, + 0.11635200679302216, + 0.12407781183719635, + 0.07465717941522598, + 0.14377343654632568, + 0.11371515691280365, + 0.06264426559209824, + 0.04765245318412781, + 0.028170999139547348 + ], + "credit_map": [ + 0.016779262572526932, + 0.02559485286474228, + 0.019481781870126724, + 0.026057133451104164, + 0.02127513475716114, + 0.012762373313307762, + 0.026543468236923218, + 0.013936564326286316, + 0.015035156160593033, + 0.0203954316675663, + 0.018480954691767693, + 0.020954251289367676, + 0.01495937816798687, + 0.015597117133438587, + 0.011505658738315105, + 0.009598145261406898, + 0.014025925658643246, + 0.014381282031536102, + 0.015922430902719498, + 0.020958373323082924, + 0.015851877629756927, + 0.013859030790627003, + 0.012294553220272064, + 0.00905853696167469, + 0.014966906979680061, + 0.01839636266231537, + 0.013602063059806824, + 0.01781417988240719, + 0.01399250514805317, + 0.01171158254146576, + 0.01789640262722969, + 0.015697814524173737, + 0.009452561847865582, + 0.008242825977504253, + 0.010300789959728718, + 0.010741758160293102, + 0.01486650574952364, + 0.006991320289671421, + 0.008927238173782825, + 0.005134176928550005, + 0.01860676519572735, + 0.014408165588974953, + 0.04003769904375076, + 0.010178588330745697, + 0.014867017976939678, + 0.026589227840304375, + 0.009563074447214603, + 0.009522899985313416, + 0.006600222550332546, + 0.045873917639255524, + 0.004023438319563866, + 0.012653473764657974, + 0.007231098599731922, + 0.027697088196873665, + 0.00512514216825366, + 0.004510780330747366, + 0.007492440287023783, + 0.011736742220818996, + 0.006814953405410051, + 0.008535150438547134, + 0.007845551706850529, + 0.007213931065052748, + 0.009417735040187836, + 0.0035877663176506758, + 0.0041797268204391, + 0.01033480279147625, + 0.013832277618348598, + 0.003817971795797348, + 0.004021763801574707, + 0.004557237960398197, + 0.0036158482544124126, + 0.0032928239088505507, + 0.002299413550645113, + 0.004012225661426783, + 0.004057679325342178, + 0.003692826721817255, + 0.004313711542636156, + 0.004002423491328955, + 0.003284416627138853, + 0.002508298959583044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.5761873126029968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137058973312378 + ], + "opposite_logprob_tokens": [ + -0.3870590031147003 + ], + "logprob_reward": -1.137058973312378, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137058973312378, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13177016377449036, + 0.09558054059743881, + 0.06136985868215561, + 0.06663559377193451, + 0.08218739926815033, + 0.33422067761421204, + 0.06220557913184166, + 0.06794901937246323, + 0.05725780874490738, + 0.04082334414124489 + ], + "credit_map": [ + 0.011708014644682407, + 0.020543714985251427, + 0.017707880586385727, + 0.029077066108584404, + 0.008194324560463428, + 0.007726683281362057, + 0.028211869299411774, + 0.008600614033639431, + 0.011190150864422321, + 0.015465857461094856, + 0.014137443155050278, + 0.01546850148588419, + 0.012366775423288345, + 0.011022841557860374, + 0.009501512162387371, + 0.006427456624805927, + 0.006073095835745335, + 0.007798114325851202, + 0.00935672502964735, + 0.009285492822527885, + 0.006504565477371216, + 0.006497538648545742, + 0.006147957406938076, + 0.009706364944577217, + 0.00790150836110115, + 0.008319396525621414, + 0.008082007057964802, + 0.012150293216109276, + 0.0087459497153759, + 0.008221322670578957, + 0.006687934510409832, + 0.006527186371386051, + 0.011232569813728333, + 0.011924549005925655, + 0.01004849188029766, + 0.016222387552261353, + 0.009595782496035099, + 0.008925490081310272, + 0.00809479970484972, + 0.006143325939774513, + 0.09004595875740051, + 0.031539663672447205, + 0.036811426281929016, + 0.02544129267334938, + 0.033495623618364334, + 0.07425786554813385, + 0.02018376626074314, + 0.02244509942829609, + 0.0041954247280955315, + 0.021009864285588264, + 0.006259202491492033, + 0.008649291470646858, + 0.004443550948053598, + 0.007301937788724899, + 0.005301876924932003, + 0.00504442909732461, + 0.009096312336623669, + 0.008774694055318832, + 0.009031062014400959, + 0.008104085922241211, + 0.008634863421320915, + 0.008710318244993687, + 0.00961922574788332, + 0.005978454370051622, + 0.003809439018368721, + 0.018778247758746147, + 0.014664921909570694, + 0.003780414117500186, + 0.005473035853356123, + 0.003924689255654812, + 0.003451415104791522, + 0.0033756450284272432, + 0.005171371623873711, + 0.006863519083708525, + 0.005034152418375015, + 0.004529336467385292, + 0.00537173543125391, + 0.00596216693520546, + 0.004508743993937969, + 0.0033823170233517885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -1.137058973312378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554189190268517 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554189190268517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554189190268517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1726473867893219, + 0.08310774713754654, + 0.11773890256881714, + 0.0590704120695591, + 0.051126398146152496, + 0.08267613500356674, + 0.21944798529148102, + 0.04620557278394699, + 0.10652094334363937, + 0.06145842745900154 + ], + "credit_map": [ + 0.01574290171265602, + 0.022572435438632965, + 0.024094827473163605, + 0.03774681314826012, + 0.01329788751900196, + 0.01074249017983675, + 0.03601038083434105, + 0.012439663521945477, + 0.008309908211231232, + 0.013644446618855, + 0.011556154116988182, + 0.013593730516731739, + 0.011621078476309776, + 0.00948208849877119, + 0.008503398858010769, + 0.006396944168955088, + 0.013795476406812668, + 0.011371861211955547, + 0.01598474755883217, + 0.01878667250275612, + 0.011531103402376175, + 0.021905597299337387, + 0.01157322246581316, + 0.012790233828127384, + 0.006389972288161516, + 0.00804566964507103, + 0.006670673377811909, + 0.011625495739281178, + 0.007858973927795887, + 0.006407373119145632, + 0.006157781928777695, + 0.005914477631449699, + 0.0070694261230528355, + 0.004988372325897217, + 0.00629355339333415, + 0.008936581201851368, + 0.008310097269713879, + 0.004808444529771805, + 0.0064363619312644005, + 0.004283566027879715, + 0.008978662081062794, + 0.00985080748796463, + 0.01518228743225336, + 0.008498834446072578, + 0.00804931577295065, + 0.02033347636461258, + 0.006387325003743172, + 0.005395438522100449, + 0.011349516920745373, + 0.08830851316452026, + 0.011211615055799484, + 0.028571536764502525, + 0.013158151879906654, + 0.045809414237737656, + 0.011117465794086456, + 0.009921793825924397, + 0.0054283058270812035, + 0.005968373268842697, + 0.00667168665677309, + 0.005566098261624575, + 0.006310525815933943, + 0.006054500583559275, + 0.005948937032371759, + 0.0042571513913571835, + 0.005841400474309921, + 0.0378982275724411, + 0.031454190611839294, + 0.0069531165063381195, + 0.008808258920907974, + 0.00607680669054389, + 0.005544224753975868, + 0.003944727126508951, + 0.008046748116612434, + 0.010883169248700142, + 0.006237916182726622, + 0.005809756461530924, + 0.00829401332885027, + 0.009987413883209229, + 0.00648781843483448, + 0.005711599253118038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.012554189190268517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement static (no panning, tilting, or zooming)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the framing an extreme close-up of a plain, unlabeled metal paint can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thick, viscous stream of vibrant crimson pigment cascade from the can and shatter into a radial explosion upon hitting the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rough, gray concrete floor illuminated by harsh, direct overhead light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038112103939056396 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.038112103939056396, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038112103939056396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28779762983322144, + 0.14314119517803192, + 0.041758038103580475, + 0.06323879957199097, + 0.0321495458483696, + 0.1839497685432434, + 0.016326481476426125, + 0.036199331283569336, + 0.08037833869457245, + 0.11506084352731705 + ], + "credit_map": [ + 0.02690359763801098, + 0.041908882558345795, + 0.03771672025322914, + 0.05862696096301079, + 0.01821434311568737, + 0.01719597540795803, + 0.06653087586164474, + 0.020700279623270035, + 0.02671757899224758, + 0.027149733155965805, + 0.021930238232016563, + 0.017667168751358986, + 0.010837040841579437, + 0.015783337876200676, + 0.011575165204703808, + 0.011480942368507385, + 0.004684651270508766, + 0.003954799845814705, + 0.0065198130905628204, + 0.007217927370220423, + 0.003816660027951002, + 0.006293818354606628, + 0.004227115772664547, + 0.005043251905590296, + 0.00621302705258131, + 0.008227162063121796, + 0.007974640466272831, + 0.007919808849692345, + 0.007687694858759642, + 0.007957438938319683, + 0.007958976551890373, + 0.009300046600401402, + 0.004701412282884121, + 0.004099791869521141, + 0.0044290474615991116, + 0.004579965956509113, + 0.00566987507045269, + 0.00270160217769444, + 0.003198162419721484, + 0.002769691403955221, + 0.05459431931376457, + 0.016459673643112183, + 0.02604370377957821, + 0.013471479527652264, + 0.012812726199626923, + 0.03568101301789284, + 0.01025866437703371, + 0.014628199860453606, + 0.0019081493373960257, + 0.003445700043812394, + 0.0024262687657028437, + 0.0018776633078232408, + 0.0014149404596537352, + 0.001438661478459835, + 0.0014996582176536322, + 0.00231543998233974, + 0.004689739551395178, + 0.0037065125070512295, + 0.003412309568375349, + 0.004189617931842804, + 0.004553514998406172, + 0.006675484590232372, + 0.003931480925530195, + 0.005040669813752174, + 0.005205114837735891, + 0.024251360446214676, + 0.022987036034464836, + 0.006774482782930136, + 0.006854317616671324, + 0.005395626183599234, + 0.005270287860184908, + 0.0036401106044650078, + 0.017177419736981392, + 0.014849209226667881, + 0.010856397449970245, + 0.011990637518465519, + 0.01506073959171772, + 0.02288930118083954, + 0.011380274780094624, + 0.010856864042580128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.038112103939056396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_tensor([5184, 5184])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_tensor([5184, 5184])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9be3d0edf83b3a8d7257f4dc30de1012c1a2e70a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_tensor([5184, 5184])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static close-up shot from a level angle focuses on an antique globe resting on the corner of a dark, solid wood desk. A soft, focused, cinematic light, like afternoon sunlight streaming from a high window, slants down, casting a precise shadow over the globe. Countless tiny dust particles are clearly visible within this beam, slowly drifting, swirling, and flickering within the light, imbuing the still image with a subtle vitality. The globe itself bears the marks of age, its stand constructed of patinaed brass and its base of heavy walnut. Its surface, the texture of yellowed parchment, bears a faded, ancient map, depicting the continents in ochre and the oceans in a pale cyan. The light precisely outlines the curved contours of the globe and forms a warm highlight on its polished surface, while enveloping the wood grain texture of the desk and the blurred outline of the bookshelf in the background in a warm, tranquil and nostalgic atmosphere. The tones of the entire scene are dominated by warm gold and deep brown, creating a quiet beauty of contemplation and time freezing.", + "seed": 13661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_5184_seed_13661_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02222125083208084, + "vlm_logprob": -0.02222125083208084, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02222125083208084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009237644262611866 + ], + "opposite_logprob_tokens": [ + -5.009237766265869 + ], + "logprob_reward": -0.009237644262611866, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009237644262611866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20501719415187836, + 0.08117365092039108, + 0.05787382647395134, + 0.06671721488237381, + 0.09102074801921844, + 0.04513319954276085, + 0.10818455368280411, + 0.2423894852399826, + 0.04107510671019554, + 0.06141503155231476 + ], + "credit_map": [ + 0.027360251173377037, + 0.018222026526927948, + 0.023069962859153748, + 0.02436935156583786, + 0.02923187054693699, + 0.01965358667075634, + 0.03253835439682007, + 0.03057178296148777, + 0.011618432588875294, + 0.005680111702531576, + 0.01656881347298622, + 0.011197543703019619, + 0.00967569649219513, + 0.008996395394206047, + 0.008052672259509563, + 0.0093839755281806, + 0.009890692308545113, + 0.003851975779980421, + 0.006782777141779661, + 0.009310645051300526, + 0.007545434404164553, + 0.008622930385172367, + 0.006173418369144201, + 0.005695947911590338, + 0.01251998171210289, + 0.004710063803941011, + 0.008601834997534752, + 0.008019880391657352, + 0.006775920279324055, + 0.010466048493981361, + 0.010189981199800968, + 0.005433493759483099, + 0.009872162714600563, + 0.0049095056019723415, + 0.0136507423594594, + 0.011693361215293407, + 0.012714400887489319, + 0.010466164909303188, + 0.019678335636854172, + 0.008036067709326744, + 0.006528597325086594, + 0.0026268381625413895, + 0.006703644525259733, + 0.004485913552343845, + 0.005048486404120922, + 0.0037315969821065664, + 0.0107493307441473, + 0.005258780904114246, + 0.02957414649426937, + 0.005562699865549803, + 0.03301996365189552, + 0.009139145724475384, + 0.005209558177739382, + 0.006931886542588472, + 0.010496956296265125, + 0.008250190876424313, + 0.05604386329650879, + 0.01610201969742775, + 0.05978449061512947, + 0.024086276069283485, + 0.01442693267017603, + 0.016900623217225075, + 0.025115150958299637, + 0.0299301128834486, + 0.010331041179597378, + 0.003279660129919648, + 0.007062628399580717, + 0.00429582828655839, + 0.0025948253460228443, + 0.0037318954709917307, + 0.004281213972717524, + 0.00549800880253315, + 0.013746371492743492, + 0.004288786556571722, + 0.006811767816543579, + 0.007728284690529108, + 0.0042046718299388885, + 0.00598430959507823, + 0.009218004532158375, + 0.00943282525986433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.009237644262611866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05506475642323494 + ], + "opposite_logprob_tokens": [ + -2.9300646781921387 + ], + "logprob_reward": -0.05506475642323494, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05506475642323494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2542571723461151, + 0.09904429316520691, + 0.0533660463988781, + 0.04396821931004524, + 0.13589240610599518, + 0.0409974530339241, + 0.07032294571399689, + 0.18231426179409027, + 0.08048415929079056, + 0.039353080093860626 + ], + "credit_map": [ + 0.0342085137963295, + 0.02566298469901085, + 0.04546007513999939, + 0.048852309584617615, + 0.019500594586133957, + 0.01754889264702797, + 0.027769966050982475, + 0.03525380417704582, + 0.014252644963562489, + 0.006066543981432915, + 0.014378122985363007, + 0.013057725504040718, + 0.009944630786776543, + 0.00968632660806179, + 0.02017848938703537, + 0.011479808948934078, + 0.008426093496382236, + 0.003906182013452053, + 0.01062918733805418, + 0.006476710084825754, + 0.005671681836247444, + 0.0057917493395507336, + 0.007570507004857063, + 0.00489393575116992, + 0.007432498969137669, + 0.00312765222042799, + 0.006508438847959042, + 0.006402566563338041, + 0.004301495384424925, + 0.007852012291550636, + 0.004388001747429371, + 0.003955553751438856, + 0.016260268166661263, + 0.008082463406026363, + 0.022706354036927223, + 0.017691493034362793, + 0.017971646040678024, + 0.014275513589382172, + 0.02705864980816841, + 0.011846012435853481, + 0.005778503604233265, + 0.0023919851519167423, + 0.008942236192524433, + 0.00515185110270977, + 0.004183115437626839, + 0.004286556039005518, + 0.005009979475289583, + 0.005253222305327654, + 0.01296838466078043, + 0.006647561676800251, + 0.01205091830343008, + 0.010063069872558117, + 0.006827249191701412, + 0.0050158388912677765, + 0.009722009301185608, + 0.007027912884950638, + 0.045674003660678864, + 0.009418586269021034, + 0.050552647560834885, + 0.017003824934363365, + 0.007907094433903694, + 0.013638606294989586, + 0.01421380415558815, + 0.02390570566058159, + 0.01881755143404007, + 0.004825258627533913, + 0.010299785993993282, + 0.00947814155369997, + 0.005686577409505844, + 0.007133371662348509, + 0.00924732256680727, + 0.01499615516513586, + 0.007563636638224125, + 0.0027421542908996344, + 0.004002941306680441, + 0.003925302531570196, + 0.0037117123138159513, + 0.0039029244799166918, + 0.008958300575613976, + 0.004546104930341244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.05506475642323494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079042106866837 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016079042106866837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079042106866837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1567699909210205, + 0.06793452054262161, + 0.037994831800460815, + 0.03985690698027611, + 0.03202230855822563, + 0.08040320873260498, + 0.06009356305003166, + 0.3342767357826233, + 0.10888703912496567, + 0.08176087588071823 + ], + "credit_map": [ + 0.02268134616315365, + 0.00794515386223793, + 0.02102060057222843, + 0.02058851718902588, + 0.010359670035541058, + 0.01848711632192135, + 0.015324130654335022, + 0.04036345332860947, + 0.010849972255527973, + 0.0042517222464084625, + 0.012632458470761776, + 0.006025961134582758, + 0.008172428235411644, + 0.005817631725221872, + 0.008556590415537357, + 0.011627756990492344, + 0.007138833403587341, + 0.0021973219700157642, + 0.0055897231213748455, + 0.004392791539430618, + 0.004846821539103985, + 0.006029807962477207, + 0.004667701665312052, + 0.003131827572360635, + 0.005534473806619644, + 0.0027951966039836407, + 0.0062780301086604595, + 0.004123841878026724, + 0.005537409335374832, + 0.0048284451477229595, + 0.005317716393619776, + 0.005441794637590647, + 0.00525434035807848, + 0.0021924490574747324, + 0.005866331979632378, + 0.004363717045634985, + 0.0023662899620831013, + 0.003053446300327778, + 0.005089898128062487, + 0.0038358373567461967, + 0.010419401340186596, + 0.00652414932847023, + 0.007506178226321936, + 0.00832499098032713, + 0.014657597988843918, + 0.008141047321259975, + 0.018206322565674782, + 0.006623517721891403, + 0.016051337122917175, + 0.0039718677289783955, + 0.013784561306238174, + 0.004783278796821833, + 0.004798371810466051, + 0.006296963430941105, + 0.00573243061080575, + 0.004674754571169615, + 0.08996505290269852, + 0.013487168587744236, + 0.10494589060544968, + 0.02345709130167961, + 0.013958333991467953, + 0.0246184803545475, + 0.025091398507356644, + 0.038753315806388855, + 0.02344735525548458, + 0.006644212640821934, + 0.01508662011474371, + 0.012801271863281727, + 0.006899181753396988, + 0.010327894240617752, + 0.013898969627916813, + 0.019781533628702164, + 0.014514569193124771, + 0.00564296031370759, + 0.009292704984545708, + 0.008742090314626694, + 0.006531897466629744, + 0.007882595993578434, + 0.020334064960479736, + 0.00881998986005783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.016079042106866837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818133518099785 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.01818133518099785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818133518099785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3651472330093384, + 0.06946249306201935, + 0.045335132628679276, + 0.029782652854919434, + 0.07614032179117203, + 0.0830228328704834, + 0.08955847471952438, + 0.12455585598945618, + 0.07159523665904999, + 0.045399755239486694 + ], + "credit_map": [ + 0.05420574173331261, + 0.036347776651382446, + 0.04409785196185112, + 0.03663082793354988, + 0.028009014204144478, + 0.03159807249903679, + 0.04380938783288002, + 0.09044857323169708, + 0.009685294702649117, + 0.0049142418429255486, + 0.011241460219025612, + 0.009917475283145905, + 0.007188586052507162, + 0.006525473203510046, + 0.010332915000617504, + 0.009657046757638454, + 0.007636804599314928, + 0.0027856917586177588, + 0.006276534404605627, + 0.006060776766389608, + 0.006026132497936487, + 0.006183568853884935, + 0.00649853702634573, + 0.0038670883513987064, + 0.004103825893253088, + 0.0020425005350261927, + 0.005323471035808325, + 0.004482215270400047, + 0.0032327515073120594, + 0.0027245059609413147, + 0.004702471196651459, + 0.0031709116883575916, + 0.010446391999721527, + 0.004903801251202822, + 0.009683516807854176, + 0.010730455629527569, + 0.010909393429756165, + 0.008763669058680534, + 0.013523678295314312, + 0.0071794139221310616, + 0.012961190193891525, + 0.005248231813311577, + 0.01459335908293724, + 0.010395419783890247, + 0.008658685721457005, + 0.006956194061785936, + 0.013090463355183601, + 0.011119283735752106, + 0.02601335011422634, + 0.004916894715279341, + 0.02457290329039097, + 0.005052335560321808, + 0.006349112838506699, + 0.009157072752714157, + 0.008580509573221207, + 0.004916294477880001, + 0.027650292962789536, + 0.006932418793439865, + 0.020391706377267838, + 0.015013904310762882, + 0.007367775309830904, + 0.009040436707437038, + 0.012353758327662945, + 0.02580556645989418, + 0.014700544066727161, + 0.004444516263902187, + 0.00868280790746212, + 0.007350721396505833, + 0.007408622652292252, + 0.006273453123867512, + 0.011816904880106449, + 0.010917662642896175, + 0.010221454314887524, + 0.003043767064809799, + 0.006186875980347395, + 0.004537655506283045, + 0.0037233326584100723, + 0.0035828426480293274, + 0.007182433735579252, + 0.006921391002833843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01818133518099785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10938301682472229, + 0.09913700819015503, + 0.11482015997171402, + 0.1102445051074028, + 0.29212692379951477, + 0.056761667132377625, + 0.016718603670597076, + 0.15263287723064423, + 0.0167964156717062, + 0.03137883171439171 + ], + "credit_map": [ + 0.0184143278747797, + 0.01325705461204052, + 0.015115954913198948, + 0.014396150596439838, + 0.007607292849570513, + 0.009024279192090034, + 0.013497928157448769, + 0.018070032820105553, + 0.019251534715294838, + 0.004557715728878975, + 0.03262108191847801, + 0.01127319410443306, + 0.005151706747710705, + 0.008293481543660164, + 0.006670954171568155, + 0.011317338794469833, + 0.016637219116091728, + 0.006950626149773598, + 0.022525107488036156, + 0.018061581999063492, + 0.012446161359548569, + 0.00966684054583311, + 0.015525764785706997, + 0.013006849214434624, + 0.014889881014823914, + 0.006923540960997343, + 0.018274834379553795, + 0.019399132579565048, + 0.013223620131611824, + 0.009782776236534119, + 0.014168000780045986, + 0.013582718558609486, + 0.045000556856393814, + 0.01571912132203579, + 0.04947013035416603, + 0.051002923399209976, + 0.019744282588362694, + 0.025510627776384354, + 0.05343334376811981, + 0.032245926558971405, + 0.010990451090037823, + 0.0032093063928186893, + 0.01308018434792757, + 0.006902407389134169, + 0.004439530894160271, + 0.004896658938378096, + 0.004896494094282389, + 0.008346633985638618, + 0.0036425995640456676, + 0.00095426628831774, + 0.0034584032837301493, + 0.0018039437709376216, + 0.0013431538827717304, + 0.0015342128463089466, + 0.002343727508559823, + 0.0016382960602641106, + 0.043767500668764114, + 0.005640041083097458, + 0.052733566612005234, + 0.009369448758661747, + 0.00557270972058177, + 0.01109405793249607, + 0.009888926520943642, + 0.014566626399755478, + 0.004026629496365786, + 0.0010120581137016416, + 0.0021144249476492405, + 0.002043290063738823, + 0.0012151802657172084, + 0.0014806208200752735, + 0.0021769965533167124, + 0.00272721447981894, + 0.0073195998556911945, + 0.0022163440007716417, + 0.0034987747203558683, + 0.0031691587064415216, + 0.0023687814828008413, + 0.0030908696353435516, + 0.005367517005652189, + 0.004347787704318762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.012543476186692715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a level angle focuses on an antique globe resting on the corner of a dark, solid wood desk. A soft, focused, cinematic light, like afternoon sunlight streaming from a high window, slants down, casting a precise shadow over the globe. Countless tiny dust particles are clearly visible within this beam, slowly drifting, swirling, and flickering within the light, imbuing the still image with a subtle vitality. The globe itself bears the marks of age, its stand constructed of patinaed brass and its base of heavy walnut. Its surface, the texture of yellowed parchment, bears a faded, ancient map, depicting the continents in ochre and the oceans in a pale cyan. The light precisely outlines the curved contours of the globe and forms a warm highlight on its polished surface, while enveloping the wood grain texture of the desk and the blurred outline of the bookshelf in the background in a warm, tranquil and nostalgic atmosphere. The tones of the entire scene are dominated by warm gold and deep brown, creating a quiet beauty of contemplation and time freezing.", + "seed": 13662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_5184_seed_13662_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01661362638697028, + "vlm_logprob": -0.01661362638697028, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01661362638697028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011839832179248333 + ], + "opposite_logprob_tokens": [ + -4.761839866638184 + ], + "logprob_reward": -0.011839832179248333, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011839832179248333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19303791224956512, + 0.07026350498199463, + 0.09517131745815277, + 0.13664455711841583, + 0.044006455689668655, + 0.05304334685206413, + 0.09513717889785767, + 0.1856078803539276, + 0.049563392996788025, + 0.07752446085214615 + ], + "credit_map": [ + 0.028208967298269272, + 0.012370173819363117, + 0.03920011594891548, + 0.01534235943108797, + 0.01041948888450861, + 0.013950568623840809, + 0.026706865057349205, + 0.04683937132358551, + 0.006938943173736334, + 0.0045415363274514675, + 0.01573433354496956, + 0.00524810841307044, + 0.005979228764772415, + 0.003583564655855298, + 0.019549118354916573, + 0.00868866965174675, + 0.010583503171801567, + 0.007962829433381557, + 0.019726300612092018, + 0.00882653333246708, + 0.012441462837159634, + 0.00611926568672061, + 0.014620596542954445, + 0.014890828169882298, + 0.017249366268515587, + 0.015271633863449097, + 0.01940253935754299, + 0.016438651829957962, + 0.014393428340554237, + 0.008383297361433506, + 0.02337023802101612, + 0.022135399281978607, + 0.005417055916041136, + 0.005404466297477484, + 0.007691064849495888, + 0.0046547711826860905, + 0.006102785002440214, + 0.002823207527399063, + 0.0050275735557079315, + 0.006885529961436987, + 0.004925904795527458, + 0.006001930683851242, + 0.009855111129581928, + 0.005684864707291126, + 0.00866694375872612, + 0.004060268867760897, + 0.006549452431499958, + 0.007298870477825403, + 0.009700750932097435, + 0.007184453308582306, + 0.01992480456829071, + 0.006972259841859341, + 0.02094085142016411, + 0.0066252113319933414, + 0.008016183972358704, + 0.015772664919495583, + 0.021348224952816963, + 0.022104687988758087, + 0.021215595304965973, + 0.03465292602777481, + 0.029165832325816154, + 0.017226679250597954, + 0.024977706372737885, + 0.014916225336492062, + 0.004355132579803467, + 0.00755669642239809, + 0.005979612935334444, + 0.0071770912036299706, + 0.005672391038388014, + 0.004683765582740307, + 0.007598701398819685, + 0.006540000904351473, + 0.0071507967077195644, + 0.0041764299385249615, + 0.007034996524453163, + 0.004036656115204096, + 0.02910870872437954, + 0.0033687890972942114, + 0.013888861052691936, + 0.008759217336773872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.011839832179248333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03818175569176674 + ], + "opposite_logprob_tokens": [ + -3.288181781768799 + ], + "logprob_reward": -0.03818175569176674, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03818175569176674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13443194329738617, + 0.0518278144299984, + 0.28333377838134766, + 0.022717537358403206, + 0.07103762775659561, + 0.04017399251461029, + 0.21456235647201538, + 0.04689093679189682, + 0.025252755731344223, + 0.10977133363485336 + ], + "credit_map": [ + 0.01119142584502697, + 0.005520329345017672, + 0.021208448335528374, + 0.007691544480621815, + 0.008006082847714424, + 0.008252490311861038, + 0.02675463818013668, + 0.04580697789788246, + 0.004089844413101673, + 0.0034036310389637947, + 0.010310091078281403, + 0.0056539373472332954, + 0.00590995978564024, + 0.002945290645584464, + 0.013105655089020729, + 0.006409403868019581, + 0.034372758120298386, + 0.018870113417506218, + 0.06480752676725388, + 0.025891464203596115, + 0.016124382615089417, + 0.01524791494011879, + 0.05452224239706993, + 0.05349735543131828, + 0.0030315816402435303, + 0.0029248581267893314, + 0.004107870627194643, + 0.0027126355562359095, + 0.0026740834582597017, + 0.0012740753591060638, + 0.003174449549987912, + 0.0028179832734167576, + 0.008400634862482548, + 0.007306101266294718, + 0.009927896782755852, + 0.006523633375763893, + 0.00914330780506134, + 0.006373480893671513, + 0.01025643851608038, + 0.013106130994856358, + 0.003899754723533988, + 0.006485524121671915, + 0.004730031825602055, + 0.005751872900873423, + 0.005150364246219397, + 0.004803675692528486, + 0.0041758338920772076, + 0.005176934879273176, + 0.015611063688993454, + 0.011660577729344368, + 0.06453700363636017, + 0.014553457498550415, + 0.0613037645816803, + 0.010850071907043457, + 0.01253325492143631, + 0.02351316623389721, + 0.0050131273455917835, + 0.005644631572067738, + 0.004743924830108881, + 0.008927890099585056, + 0.0057539199478924274, + 0.005577889271080494, + 0.006760877091437578, + 0.004468675237149, + 0.002684393199160695, + 0.004315864760428667, + 0.0031955025624483824, + 0.00328904134221375, + 0.0033371755853295326, + 0.002469221828505397, + 0.0030196174047887325, + 0.0029419411439448595, + 0.004706248175352812, + 0.004460114520043135, + 0.005185817833989859, + 0.004035168327391148, + 0.04517264664173126, + 0.004774653818458319, + 0.036829665303230286, + 0.004607013892382383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03818175569176674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210931025445461 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210931025445461, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210931025445461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09074588865041733, + 0.09758210182189941, + 0.06371073424816132, + 0.0503978431224823, + 0.030645053833723068, + 0.10376951098442078, + 0.40748026967048645, + 0.028205711394548416, + 0.05333999544382095, + 0.0741228312253952 + ], + "credit_map": [ + 0.017243582755327225, + 0.008364484645426273, + 0.010527175851166248, + 0.009946459904313087, + 0.008897621184587479, + 0.007416462991386652, + 0.015255539678037167, + 0.013094562105834484, + 0.010011516511440277, + 0.0066506583243608475, + 0.02112247236073017, + 0.009166873060166836, + 0.007593068294227123, + 0.005650818347930908, + 0.027257626876235008, + 0.010129066184163094, + 0.006657955702394247, + 0.006789804436266422, + 0.016123438253998756, + 0.005457771476358175, + 0.0072916001081466675, + 0.0032860860228538513, + 0.010545962490141392, + 0.007558113429695368, + 0.005459318868815899, + 0.0046933796256780624, + 0.006630939897149801, + 0.007382395211607218, + 0.006434727925807238, + 0.0033935518004000187, + 0.007605062797665596, + 0.008798466064035892, + 0.0030590256210416555, + 0.004096921067684889, + 0.0043228184804320335, + 0.0025052011478692293, + 0.004356766119599342, + 0.003259639022871852, + 0.0045333621092140675, + 0.004511320032179356, + 0.010276841931045055, + 0.007956509478390217, + 0.009820987470448017, + 0.00555591844022274, + 0.029992934316396713, + 0.006210003513842821, + 0.006171474698930979, + 0.027784839272499084, + 0.028189897537231445, + 0.018864603713154793, + 0.13356854021549225, + 0.026885753497481346, + 0.12504233419895172, + 0.017279407009482384, + 0.01834951899945736, + 0.03930022940039635, + 0.0033021881245076656, + 0.003590763546526432, + 0.004400485195219517, + 0.003435445949435234, + 0.0047611454501748085, + 0.002788483863696456, + 0.0026934680063277483, + 0.0032337342854589224, + 0.003124202834442258, + 0.005975781008601189, + 0.0076538450084626675, + 0.006046820431947708, + 0.006915008183568716, + 0.006689209956675768, + 0.010074972175061703, + 0.006860159803181887, + 0.0033238388132303953, + 0.0036349534057080746, + 0.005891356151551008, + 0.004577795509248972, + 0.022187501192092896, + 0.0050522745586931705, + 0.020560838282108307, + 0.008894274942576885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.014210931025445461 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12974749505519867, + 0.047576967626810074, + 0.125498965382576, + 0.023025188595056534, + 0.02427816390991211, + 0.22881579399108887, + 0.18931274116039276, + 0.03770028054714203, + 0.07593341171741486, + 0.1181110367178917 + ], + "credit_map": [ + 0.02050272561609745, + 0.011046824976801872, + 0.02402174472808838, + 0.015004802495241165, + 0.008537722751498222, + 0.008284216746687889, + 0.019716886803507805, + 0.022632574662566185, + 0.003849891247227788, + 0.0018878972623497248, + 0.011983886361122131, + 0.002394218696281314, + 0.0025834278203547, + 0.0020429901778697968, + 0.017152462154626846, + 0.0056821913458406925, + 0.01523797120898962, + 0.008213293738663197, + 0.03447677567601204, + 0.009291831403970718, + 0.006155538372695446, + 0.0058013140223920345, + 0.0236005000770092, + 0.022721732035279274, + 0.0035783671773970127, + 0.00291651114821434, + 0.003445145906880498, + 0.0028195942286401987, + 0.0031350390054285526, + 0.0018803924322128296, + 0.002185598947107792, + 0.003064539283514023, + 0.002634536474943161, + 0.002308471594005823, + 0.004237090703099966, + 0.002291691955178976, + 0.002932608826085925, + 0.001439634128473699, + 0.0032401634380221367, + 0.005193965043872595, + 0.017720572650432587, + 0.036388784646987915, + 0.03117891028523445, + 0.02168310061097145, + 0.03848631680011749, + 0.009003836661577225, + 0.017474345862865448, + 0.0568799264729023, + 0.013293469324707985, + 0.008278513327240944, + 0.0637211799621582, + 0.012312149628996849, + 0.05923829600214958, + 0.007115153595805168, + 0.007704093120992184, + 0.01764989085495472, + 0.004091581329703331, + 0.00362773728556931, + 0.004212772939354181, + 0.004460127092897892, + 0.005851962603628635, + 0.0066926670260727406, + 0.004581291228532791, + 0.004182137548923492, + 0.0071984510868787766, + 0.010364257730543613, + 0.008839759975671768, + 0.013317400589585304, + 0.010685877874493599, + 0.00481863459572196, + 0.013881023041903973, + 0.00682800030335784, + 0.010646265000104904, + 0.007114265114068985, + 0.020727159455418587, + 0.007278820034116507, + 0.037171829491853714, + 0.004852981306612492, + 0.017920831218361855, + 0.01239889394491911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643016029149294 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643016029149294, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643016029149294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18988646566867828, + 0.3145700991153717, + 0.0553697794675827, + 0.09743888676166534, + 0.01883964240550995, + 0.04048183187842369, + 0.1778579205274582, + 0.01780560612678528, + 0.03207932412624359, + 0.05567040666937828 + ], + "credit_map": [ + 0.015327257104218006, + 0.010456562042236328, + 0.04865819215774536, + 0.012584543786942959, + 0.011590748094022274, + 0.011292374692857265, + 0.02524554170668125, + 0.05473124980926514, + 0.034071918576955795, + 0.015469747595489025, + 0.06420578062534332, + 0.023691212758421898, + 0.01794484630227089, + 0.019460396841168404, + 0.10192486643791199, + 0.0378013551235199, + 0.005940236616879702, + 0.00402303459122777, + 0.01391618512570858, + 0.0038248493801802397, + 0.003851918736472726, + 0.0029204650782048702, + 0.013215149752795696, + 0.007677942980080843, + 0.010024136863648891, + 0.015238353051245213, + 0.013366054743528366, + 0.008542773313820362, + 0.012026223354041576, + 0.005828280001878738, + 0.016528479754924774, + 0.015884583815932274, + 0.001970186596736312, + 0.0027525518089532852, + 0.0025386561173945665, + 0.0019198807422071695, + 0.0028481590561568737, + 0.001725518610328436, + 0.002571114804595709, + 0.002513573970645666, + 0.0038078036159276962, + 0.003657638095319271, + 0.004359713755548, + 0.0023651490919291973, + 0.010483096353709698, + 0.0022953615989536047, + 0.0021733648609369993, + 0.011339700780808926, + 0.020380251109600067, + 0.016486546024680138, + 0.01981542259454727, + 0.013870807364583015, + 0.027362728491425514, + 0.01753006875514984, + 0.027919471263885498, + 0.034492626786231995, + 0.0023059328086674213, + 0.0019617462530732155, + 0.0028714791405946016, + 0.0019172972533851862, + 0.003643520874902606, + 0.001593095134012401, + 0.002129760105162859, + 0.0013827730435878038, + 0.002104892861098051, + 0.0029729348607361317, + 0.0041775754652917385, + 0.008207673206925392, + 0.004017357248812914, + 0.0018982895417138934, + 0.00585588114336133, + 0.002844718284904957, + 0.004500926937907934, + 0.002323375316336751, + 0.0041449228301644325, + 0.0032315850257873535, + 0.020846622064709663, + 0.0023115319199860096, + 0.013140149414539337, + 0.005171291064471006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.004643016029149294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a level angle focuses on an antique globe resting on the corner of a dark, solid wood desk. A soft, focused, cinematic light, like afternoon sunlight streaming from a high window, slants down, casting a precise shadow over the globe. Countless tiny dust particles are clearly visible within this beam, slowly drifting, swirling, and flickering within the light, imbuing the still image with a subtle vitality. The globe itself bears the marks of age, its stand constructed of patinaed brass and its base of heavy walnut. Its surface, the texture of yellowed parchment, bears a faded, ancient map, depicting the continents in ochre and the oceans in a pale cyan. The light precisely outlines the curved contours of the globe and forms a warm highlight on its polished surface, while enveloping the wood grain texture of the desk and the blurred outline of the bookshelf in the background in a warm, tranquil and nostalgic atmosphere. The tones of the entire scene are dominated by warm gold and deep brown, creating a quiet beauty of contemplation and time freezing.", + "seed": 13663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_5184_seed_13663_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01888081943616271, + "vlm_logprob": -0.01888081943616271, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01888081943616271, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007833120413124561 + ], + "opposite_logprob_tokens": [ + -5.382833003997803 + ], + "logprob_reward": -0.007833120413124561, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007833120413124561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2621676027774811, + 0.08330650627613068, + 0.07612704485654831, + 0.08092952519655228, + 0.07226300239562988, + 0.12410998344421387, + 0.052634820342063904, + 0.10743125528097153, + 0.10043058544397354, + 0.04059964418411255 + ], + "credit_map": [ + 0.04178238660097122, + 0.02226564846932888, + 0.020215673372149467, + 0.021577198058366776, + 0.03266166150569916, + 0.0387343131005764, + 0.033494964241981506, + 0.051435768604278564, + 0.014136766083538532, + 0.008913910947740078, + 0.006529225502163172, + 0.007275979034602642, + 0.011594392359256744, + 0.011362377554178238, + 0.011167662218213081, + 0.012326190248131752, + 0.015920206904411316, + 0.007616955786943436, + 0.005711204372346401, + 0.006638688035309315, + 0.012465879321098328, + 0.01010369323194027, + 0.01001245528459549, + 0.007657969370484352, + 0.013186353258788586, + 0.009872518479824066, + 0.004727008752524853, + 0.006855196785181761, + 0.013534569181501865, + 0.009990351274609566, + 0.01204744540154934, + 0.010716081596910954, + 0.01460917666554451, + 0.0054167346097528934, + 0.0059472923167049885, + 0.005722478032112122, + 0.007174252532422543, + 0.010584698989987373, + 0.009113838896155357, + 0.01369453314691782, + 0.022819554433226585, + 0.014213494025170803, + 0.0045491415075957775, + 0.010073167271912098, + 0.019038720056414604, + 0.01713472791016102, + 0.01853644661605358, + 0.01774473302066326, + 0.009200659580528736, + 0.004281103610992432, + 0.0037624696269631386, + 0.003542422316968441, + 0.009614065289497375, + 0.007475029211491346, + 0.005877681076526642, + 0.008881389163434505, + 0.01713886670768261, + 0.012812414206564426, + 0.006725477520376444, + 0.00892082042992115, + 0.015453604981303215, + 0.014848661608994007, + 0.013184092938899994, + 0.01834731549024582, + 0.01595526747405529, + 0.011223047971725464, + 0.005607380531728268, + 0.010982085950672626, + 0.015418143011629581, + 0.013351622968912125, + 0.012653113342821598, + 0.015239923261106014, + 0.0056203315034508705, + 0.003098999848589301, + 0.0028501413762569427, + 0.0034792388323694468, + 0.005401272792369127, + 0.0057070511393249035, + 0.00857208576053381, + 0.005870521999895573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.007833120413124561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04312505945563316 + ], + "opposite_logprob_tokens": [ + -3.1681251525878906 + ], + "logprob_reward": -0.04312505945563316, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04312505945563316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22595155239105225, + 0.11808493733406067, + 0.1191333681344986, + 0.07698121666908264, + 0.05445502698421478, + 0.0803346037864685, + 0.13523627817630768, + 0.03304664418101311, + 0.11895730346441269, + 0.03781905397772789 + ], + "credit_map": [ + 0.03504301607608795, + 0.02786119654774666, + 0.017349286004900932, + 0.017674902454018593, + 0.024376915767788887, + 0.02951866202056408, + 0.0392041839659214, + 0.03492339327931404, + 0.01643192209303379, + 0.014012620784342289, + 0.009723090566694736, + 0.010611621662974358, + 0.016551468521356583, + 0.01591193675994873, + 0.01628153957426548, + 0.0185607448220253, + 0.01358054019510746, + 0.012750604189932346, + 0.006737160962074995, + 0.008645772002637386, + 0.02296002395451069, + 0.01667661778628826, + 0.020471546798944473, + 0.01731109991669655, + 0.01176915131509304, + 0.010302828624844551, + 0.0041915480978786945, + 0.006764864549040794, + 0.01010240986943245, + 0.010371885262429714, + 0.012040079571306705, + 0.011438448913395405, + 0.007059270050376654, + 0.00388929876498878, + 0.003677332540974021, + 0.004748365841805935, + 0.00891999900341034, + 0.008220355957746506, + 0.009604065679013729, + 0.008336338214576244, + 0.012821342796087265, + 0.008956743404269218, + 0.006046874914318323, + 0.0074059972539544106, + 0.011399406008422375, + 0.010390320792794228, + 0.011310218833386898, + 0.012003698386251926, + 0.018226858228445053, + 0.018194099888205528, + 0.0056386771611869335, + 0.013249332085251808, + 0.020913412794470787, + 0.016720006242394447, + 0.026329265907406807, + 0.015964621677994728, + 0.005249358247965574, + 0.0025181907694786787, + 0.002273599151521921, + 0.0038149605970829725, + 0.004841035231947899, + 0.004137436393648386, + 0.005831705406308174, + 0.004380358848720789, + 0.01835116371512413, + 0.013675611466169357, + 0.007868858985602856, + 0.012392604723572731, + 0.015009891241788864, + 0.0158279649913311, + 0.017674727365374565, + 0.018156487494707108, + 0.0064104883931577206, + 0.003766680834814906, + 0.002619504928588867, + 0.0035124311689287424, + 0.004925687797367573, + 0.004780636169016361, + 0.006585379596799612, + 0.005218245554715395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.04312505945563316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071300953626633 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071300953626633, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071300953626633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941165268421173, + 0.15419495105743408, + 0.15726274251937866, + 0.05986500158905983, + 0.053434524685144424, + 0.05636981502175331, + 0.0418144091963768, + 0.059364303946495056, + 0.09451676905155182, + 0.029061011970043182 + ], + "credit_map": [ + 0.050319474190473557, + 0.03162769600749016, + 0.021413275972008705, + 0.020331580191850662, + 0.04863343387842178, + 0.041809648275375366, + 0.03801451250910759, + 0.04196692258119583, + 0.017224427312612534, + 0.013663276098668575, + 0.010294228792190552, + 0.011553012765944004, + 0.03372839465737343, + 0.0235730092972517, + 0.023200126364827156, + 0.02095847763121128, + 0.015337087213993073, + 0.013849480077624321, + 0.00914229266345501, + 0.010259557515382767, + 0.038628917187452316, + 0.024087047204375267, + 0.021519431844353676, + 0.024438917636871338, + 0.010432149283587933, + 0.0052339257672429085, + 0.0031263763085007668, + 0.005736798979341984, + 0.00732742203399539, + 0.007966862060129642, + 0.01192791759967804, + 0.00811354722827673, + 0.009868647903203964, + 0.005734443198889494, + 0.00328251114115119, + 0.0043669817969202995, + 0.007696608081459999, + 0.007381802424788475, + 0.007850762456655502, + 0.007252764422446489, + 0.007541324943304062, + 0.008636840619146824, + 0.0032358632888644934, + 0.004719703458249569, + 0.007999781519174576, + 0.0073067923076450825, + 0.008906295523047447, + 0.0080232173204422, + 0.006047156639397144, + 0.004676198121160269, + 0.0019818206783384085, + 0.0036420354153960943, + 0.006693585310131311, + 0.006016464438289404, + 0.006750647444278002, + 0.006006499752402306, + 0.013209589757025242, + 0.0068719335831701756, + 0.0026865247637033463, + 0.0061775678768754005, + 0.006489037070423365, + 0.0071713593788445, + 0.009159360080957413, + 0.007598928175866604, + 0.014950541779398918, + 0.011701824143528938, + 0.0057013798505067825, + 0.009885442443192005, + 0.012207169085741043, + 0.012319891713559628, + 0.013876227661967278, + 0.013874298892915249, + 0.004565506242215633, + 0.002414725022390485, + 0.002278436440974474, + 0.002932993695139885, + 0.004718806128948927, + 0.00375801301561296, + 0.004812368657439947, + 0.0035801625344902277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016071300953626633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327864058315754 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327864058315754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327864058315754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2228022813796997, + 0.07499848306179047, + 0.05137484893202782, + 0.097132109105587, + 0.1113070398569107, + 0.0904245525598526, + 0.0536959245800972, + 0.11706744879484177, + 0.11507350206375122, + 0.06612375378608704 + ], + "credit_map": [ + 0.03754781186580658, + 0.018679486587643623, + 0.022914033383131027, + 0.02268127351999283, + 0.02539053000509739, + 0.029119443148374557, + 0.0319642648100853, + 0.034505460411310196, + 0.013998931273818016, + 0.005912312772125006, + 0.00461397971957922, + 0.006691258400678635, + 0.009783084504306316, + 0.01083314884454012, + 0.011704451404511929, + 0.011461328715085983, + 0.0050520640797913074, + 0.005262073129415512, + 0.00344536523334682, + 0.004076682962477207, + 0.00881610531359911, + 0.0077278935350477695, + 0.008643489331007004, + 0.008351178839802742, + 0.018716011196374893, + 0.007060641422867775, + 0.010735377669334412, + 0.008244959637522697, + 0.006381084676831961, + 0.014816473238170147, + 0.012701989151537418, + 0.01847558654844761, + 0.021433714777231216, + 0.00820018071681261, + 0.011471317149698734, + 0.011555947363376617, + 0.008292832411825657, + 0.015492228791117668, + 0.014783734455704689, + 0.020077094435691833, + 0.018757706508040428, + 0.007473176345229149, + 0.008806253783404827, + 0.00788600742816925, + 0.007316840812563896, + 0.012289869599044323, + 0.012203024700284004, + 0.01569167710840702, + 0.006998786702752113, + 0.005811459384858608, + 0.002606950933113694, + 0.00512758269906044, + 0.007450568024069071, + 0.00755090219900012, + 0.010734559036791325, + 0.007415116764605045, + 0.029634958133101463, + 0.007439531851559877, + 0.010246389545500278, + 0.00937966350466013, + 0.010435415431857109, + 0.01576019823551178, + 0.015275243669748306, + 0.018896056339144707, + 0.017042189836502075, + 0.013817496597766876, + 0.005834278184920549, + 0.012288998812437057, + 0.015218667685985565, + 0.014876248314976692, + 0.020518044009804726, + 0.015477597713470459, + 0.0142978485673666, + 0.004323615692555904, + 0.005616739392280579, + 0.0064705731347203255, + 0.005988581571727991, + 0.008993417955935001, + 0.009366985410451889, + 0.011066003702580929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02327864058315754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095975775271654 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095975775271654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095975775271654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14995615184307098, + 0.12535080313682556, + 0.08235301822423935, + 0.09881513565778732, + 0.07694189250469208, + 0.24677430093288422, + 0.06479411572217941, + 0.04426174238324165, + 0.07228539884090424, + 0.038467392325401306 + ], + "credit_map": [ + 0.026015006005764008, + 0.013622297905385494, + 0.012466986663639545, + 0.012554971501231194, + 0.01424630917608738, + 0.023431282490491867, + 0.021248523145914078, + 0.026370778679847717, + 0.021659955382347107, + 0.010892623104155064, + 0.009917990304529667, + 0.009661355055868626, + 0.013693071901798248, + 0.019783269613981247, + 0.018504083156585693, + 0.021238455548882484, + 0.011844766326248646, + 0.0070743681862950325, + 0.006895799655467272, + 0.0069479201920330524, + 0.009913966991007328, + 0.012216448783874512, + 0.016594357788562775, + 0.010865393094718456, + 0.018994444981217384, + 0.007576921489089727, + 0.010969830676913261, + 0.009475004859268665, + 0.0063645970076322556, + 0.015308797359466553, + 0.011626600287854671, + 0.018498942255973816, + 0.014040187932550907, + 0.006278415210545063, + 0.005710236262530088, + 0.0068494463339447975, + 0.013556811958551407, + 0.00951364729553461, + 0.010700391605496407, + 0.010292759165167809, + 0.06047343462705612, + 0.01626032032072544, + 0.018655389547348022, + 0.01899016462266445, + 0.032000645995140076, + 0.034212421625852585, + 0.03029349073767662, + 0.03588844835758209, + 0.010515520349144936, + 0.009149393998086452, + 0.0024566161446273327, + 0.005909447558224201, + 0.00954766757786274, + 0.008481291122734547, + 0.00983433797955513, + 0.008899841457605362, + 0.01078522764146328, + 0.004186563193798065, + 0.0026121530681848526, + 0.0034040063619613647, + 0.005900741089135408, + 0.005526903551071882, + 0.005533868446946144, + 0.006312277168035507, + 0.013235523365437984, + 0.008744820952415466, + 0.003951659891754389, + 0.006401967257261276, + 0.008684461005032063, + 0.009583966806530952, + 0.011142692528665066, + 0.010540306568145752, + 0.007287390064448118, + 0.0035684066824615, + 0.002195365959778428, + 0.0032015761826187372, + 0.006010927259922028, + 0.0046995338052511215, + 0.006380707025527954, + 0.005123483017086983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.004095975775271654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a level angle focuses on an antique globe resting on the corner of a dark, solid wood desk. A soft, focused, cinematic light, like afternoon sunlight streaming from a high window, slants down, casting a precise shadow over the globe. Countless tiny dust particles are clearly visible within this beam, slowly drifting, swirling, and flickering within the light, imbuing the still image with a subtle vitality. The globe itself bears the marks of age, its stand constructed of patinaed brass and its base of heavy walnut. Its surface, the texture of yellowed parchment, bears a faded, ancient map, depicting the continents in ochre and the oceans in a pale cyan. The light precisely outlines the curved contours of the globe and forms a warm highlight on its polished surface, while enveloping the wood grain texture of the desk and the blurred outline of the bookshelf in the background in a warm, tranquil and nostalgic atmosphere. The tones of the entire scene are dominated by warm gold and deep brown, creating a quiet beauty of contemplation and time freezing.", + "seed": 13664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_8/index_5184_seed_13664_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019835113920271395, + "vlm_logprob": -0.019835113920271395, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.019835113920271395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014710296876728535 + ], + "opposite_logprob_tokens": [ + -4.514710426330566 + ], + "logprob_reward": -0.014710296876728535, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014710296876728535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14463269710540771, + 0.15287521481513977, + 0.0749572142958641, + 0.07567234337329865, + 0.03839827701449394, + 0.18403343856334686, + 0.12567105889320374, + 0.05984179303050041, + 0.09372314810752869, + 0.050194818526506424 + ], + "credit_map": [ + 0.018153415992856026, + 0.02738029509782791, + 0.02176443487405777, + 0.016653534024953842, + 0.014009448699653149, + 0.014386524446308613, + 0.015889737755060196, + 0.016395319253206253, + 0.037034302949905396, + 0.013398156501352787, + 0.011945364996790886, + 0.02048279531300068, + 0.011213015764951706, + 0.025139473378658295, + 0.01311352476477623, + 0.020548569038510323, + 0.0075872414745390415, + 0.010818902403116226, + 0.00798046588897705, + 0.014676367864012718, + 0.007649883162230253, + 0.008143198676407337, + 0.00873029325157404, + 0.009370865300297737, + 0.01650662161409855, + 0.009268288500607014, + 0.01314927451312542, + 0.004554568324238062, + 0.005644968245178461, + 0.008778072893619537, + 0.005485981237143278, + 0.012284564785659313, + 0.005397244822233915, + 0.00658465176820755, + 0.004005473107099533, + 0.004423223901540041, + 0.004765656311064959, + 0.0038055540062487125, + 0.0040205251425504684, + 0.005395949352532625, + 0.059166427701711655, + 0.01594238542020321, + 0.011294777505099773, + 0.008414765819907188, + 0.012062672525644302, + 0.014928613789379597, + 0.006746950093656778, + 0.0554768405854702, + 0.012206459417939186, + 0.013440779410302639, + 0.0433097779750824, + 0.009927800856530666, + 0.009715665131807327, + 0.0053138029761612415, + 0.006032323930412531, + 0.02572445012629032, + 0.005481651984155178, + 0.012149016372859478, + 0.013775207102298737, + 0.0036574210971593857, + 0.0062035261653363705, + 0.00564105249941349, + 0.0037510737311095, + 0.009182842448353767, + 0.008146385662257671, + 0.0133379390463233, + 0.020301612094044685, + 0.008923608809709549, + 0.01116958912461996, + 0.008551626466214657, + 0.0070292046293616295, + 0.01626317948102951, + 0.004255956504493952, + 0.009184449911117554, + 0.006812065839767456, + 0.005270577035844326, + 0.006751552689820528, + 0.0038795978762209415, + 0.005090720020234585, + 0.00894989725202322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.014710296876728535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048815663903951645 + ], + "opposite_logprob_tokens": [ + -3.0488157272338867 + ], + "logprob_reward": -0.048815663903951645, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048815663903951645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1312353014945984, + 0.10343535244464874, + 0.07835791260004044, + 0.0639728456735611, + 0.07233825325965881, + 0.24391695857048035, + 0.07548970729112625, + 0.06707005947828293, + 0.031754862517118454, + 0.13242870569229126 + ], + "credit_map": [ + 0.01332793664187193, + 0.02419068105518818, + 0.020203495398163795, + 0.011899757198989391, + 0.013897167518734932, + 0.01839420571923256, + 0.012974102050065994, + 0.016347957774996758, + 0.01926480606198311, + 0.011966783553361893, + 0.00859341025352478, + 0.011877212673425674, + 0.008404654450714588, + 0.01881280355155468, + 0.008632592856884003, + 0.01588308997452259, + 0.009473767131567001, + 0.009300325997173786, + 0.008159725926816463, + 0.012850122526288033, + 0.007266069762408733, + 0.007717625703662634, + 0.012702223844826221, + 0.010888054035604, + 0.011732567101716995, + 0.010579419322311878, + 0.013377766124904156, + 0.005815509706735611, + 0.006364766508340836, + 0.00467338040471077, + 0.00554433511570096, + 0.005885099060833454, + 0.00787788350135088, + 0.009517103433609009, + 0.014957468956708908, + 0.007561984937638044, + 0.008036056533455849, + 0.008246044628322124, + 0.00818213913589716, + 0.007959570735692978, + 0.0729542151093483, + 0.024508368223905563, + 0.011518983170390129, + 0.01542611327022314, + 0.01568399742245674, + 0.01638518087565899, + 0.010840266942977905, + 0.07659982889890671, + 0.014944867230951786, + 0.01261644996702671, + 0.013808784075081348, + 0.007083139382302761, + 0.008255965076386929, + 0.006173133384436369, + 0.0046087526716291904, + 0.007998618297278881, + 0.005089580547064543, + 0.0071296109817922115, + 0.01855883188545704, + 0.0034729819744825363, + 0.005935957655310631, + 0.006666213274002075, + 0.00476711243391037, + 0.015449771657586098, + 0.0032479658257216215, + 0.00508623942732811, + 0.0039400020614266396, + 0.004407510161399841, + 0.0030067460611462593, + 0.0033147050999104977, + 0.002585184993222356, + 0.006166509352624416, + 0.01074779313057661, + 0.027634955942630768, + 0.02452525869011879, + 0.017455602064728737, + 0.0157683864235878, + 0.007883980870246887, + 0.010334165766835213, + 0.018078571185469627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.048815663903951645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608525961637497 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608525961637497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608525961637497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08709697425365448, + 0.07963769882917404, + 0.017235437408089638, + 0.037231411784887314, + 0.04902714490890503, + 0.28720512986183167, + 0.10122711956501007, + 0.11001843214035034, + 0.09617134928703308, + 0.1351493000984192 + ], + "credit_map": [ + 0.009726454503834248, + 0.017101887613534927, + 0.011211221106350422, + 0.011210220865905285, + 0.010177556425333023, + 0.008135809563100338, + 0.009453359059989452, + 0.01008046418428421, + 0.008007959462702274, + 0.008203377015888691, + 0.008756027556955814, + 0.007462864276021719, + 0.005369093269109726, + 0.0159022007137537, + 0.005506217945367098, + 0.020429963245987892, + 0.001797595527023077, + 0.0033340835943818092, + 0.00230656610801816, + 0.0022443034686148167, + 0.0015430876519531012, + 0.0016825109487399459, + 0.001975318882614374, + 0.0023519713431596756, + 0.004185958299785852, + 0.008132418617606163, + 0.004373470786958933, + 0.005781251937150955, + 0.0033179577440023422, + 0.0033699481282383204, + 0.0032706474885344505, + 0.004799757152795792, + 0.008388443849980831, + 0.006699377205222845, + 0.007773829624056816, + 0.0038211122155189514, + 0.0036609095986932516, + 0.004989633336663246, + 0.0037549727130681276, + 0.009938866831362247, + 0.08542526513338089, + 0.024976087734103203, + 0.01950748823583126, + 0.01734119839966297, + 0.019649969413876534, + 0.025780873373150826, + 0.015264359302818775, + 0.07925989478826523, + 0.008170606568455696, + 0.008530884049832821, + 0.03621343895792961, + 0.006223502103239298, + 0.007841898128390312, + 0.00609495909884572, + 0.003979702014476061, + 0.024172132834792137, + 0.009278159588575363, + 0.02501768432557583, + 0.023750202730298042, + 0.008512476459145546, + 0.011431317776441574, + 0.010951678268611431, + 0.007630218286067247, + 0.013446693308651447, + 0.008390311151742935, + 0.012669097632169724, + 0.02017769031226635, + 0.00922065507620573, + 0.009445331990718842, + 0.009095889516174793, + 0.00808554794639349, + 0.01908683218061924, + 0.01184158306568861, + 0.03273259103298187, + 0.021452059969305992, + 0.01431318186223507, + 0.017581084743142128, + 0.008005121722817421, + 0.009619709104299545, + 0.019603971391916275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.01608525961637497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779518470168114 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779518470168114, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779518470168114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2886074483394623, + 0.05500321462750435, + 0.030825184658169746, + 0.017484499141573906, + 0.016698261722922325, + 0.49529343843460083, + 0.0317285992205143, + 0.0289742574095726, + 0.021436545997858047, + 0.013948576524853706 + ], + "credit_map": [ + 0.029108671471476555, + 0.04448137432336807, + 0.037573445588350296, + 0.040308207273483276, + 0.033841993659734726, + 0.035037487745285034, + 0.037462349981069565, + 0.0307939276099205, + 0.0077803051099181175, + 0.006084265187382698, + 0.005556710530072451, + 0.006251677870750427, + 0.004803894553333521, + 0.008812295272946358, + 0.004666443448513746, + 0.011047623120248318, + 0.004078085999935865, + 0.004697284195572138, + 0.00353839248418808, + 0.004558845888823271, + 0.0037391039077192545, + 0.002965725027024746, + 0.003919696435332298, + 0.0033280490897595882, + 0.0019478419562801719, + 0.002700317883864045, + 0.001970878569409251, + 0.002568574622273445, + 0.0012597352033481002, + 0.0028689811006188393, + 0.0016318310517817736, + 0.002536339918151498, + 0.0017437349306419492, + 0.001886929851025343, + 0.0033919953275471926, + 0.00156083470210433, + 0.0011364342644810677, + 0.002025572583079338, + 0.001165342633612454, + 0.0037874174304306507, + 0.1589573472738266, + 0.04167524352669716, + 0.027835892513394356, + 0.024256791919469833, + 0.03096914105117321, + 0.03828448802232742, + 0.018694348633289337, + 0.15462017059326172, + 0.005273941904306412, + 0.0034087139647454023, + 0.0049886927008628845, + 0.002487529069185257, + 0.0036370926536619663, + 0.002420480363070965, + 0.0018688420532271266, + 0.0076433042995631695, + 0.0024468956980854273, + 0.005552257876843214, + 0.007127025630325079, + 0.0019375835545361042, + 0.00290172197856009, + 0.0027842700947076082, + 0.0017499945824965835, + 0.0044745090417563915, + 0.0020930306054651737, + 0.0038229539059102535, + 0.003033257555216551, + 0.0023311658296734095, + 0.0019746499601751566, + 0.002248480450361967, + 0.0018214347073808312, + 0.004111573100090027, + 0.002441568300127983, + 0.002728734165430069, + 0.0020615921821445227, + 0.001141078770160675, + 0.0018465303583070636, + 0.0012185494415462017, + 0.0007001397316344082, + 0.0018103834008798003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.009779518470168114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up of an antique globe resting on a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the globe's stand made of patinated brass and its base made of heavy walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the globe surface show yellowed parchment texture with a faded map featuring ochre continents and pale cyan oceans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft, focused light source casting a precise shadow and illuminating visible dust particles drifting within the beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by warm gold and deep brown tones with a blurred bookshelf in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978483073413372 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.00978483073413372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978483073413372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15113621950149536, + 0.10982666909694672, + 0.16779187321662903, + 0.11245644092559814, + 0.07881347835063934, + 0.038325581699609756, + 0.13215284049510956, + 0.10638695955276489, + 0.06188911199569702, + 0.04122089967131615 + ], + "credit_map": [ + 0.017487941309809685, + 0.015217029489576817, + 0.02146078832447529, + 0.014610291458666325, + 0.014160316437482834, + 0.04053160548210144, + 0.013250447809696198, + 0.0144177982583642, + 0.02374369278550148, + 0.010359722189605236, + 0.016189079731702805, + 0.017243903130292892, + 0.006136689335107803, + 0.01069358829408884, + 0.011791071854531765, + 0.013668915256857872, + 0.022248679772019386, + 0.015230708755552769, + 0.017008211463689804, + 0.018561234697699547, + 0.014564257115125656, + 0.026695800945162773, + 0.0195189006626606, + 0.033964067697525024, + 0.019402338191866875, + 0.017496507614850998, + 0.021271754056215286, + 0.011054105125367641, + 0.008334260433912277, + 0.012108677066862583, + 0.0090004438534379, + 0.013788359239697456, + 0.011141104623675346, + 0.008180942386388779, + 0.01397273875772953, + 0.00837891735136509, + 0.006556169129908085, + 0.011585098691284657, + 0.0071433293633162975, + 0.011855174787342548, + 0.0032664574682712555, + 0.004690646659582853, + 0.005055658519268036, + 0.003936505876481533, + 0.0040663923136889935, + 0.007167283445596695, + 0.004600454121828079, + 0.005542183294892311, + 0.04532953351736069, + 0.01060481183230877, + 0.010877111926674843, + 0.008191058412194252, + 0.011128757148981094, + 0.007744199596345425, + 0.005331153515726328, + 0.03294621407985687, + 0.008312230929732323, + 0.011653978377580643, + 0.026826949790120125, + 0.005280761513859034, + 0.010391011834144592, + 0.013643711805343628, + 0.006321127060800791, + 0.02395719476044178, + 0.005506077781319618, + 0.007444228045642376, + 0.012739807367324829, + 0.006160299759358168, + 0.005808188114315271, + 0.005585450679063797, + 0.005223901942372322, + 0.01342116016894579, + 0.0045401654206216335, + 0.008158261887729168, + 0.005014265421777964, + 0.004265126306563616, + 0.0053345332853496075, + 0.002786125522106886, + 0.003833381226286292, + 0.007289041765034199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.00978483073413372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_tensor([3540, 3540])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_tensor([3540, 3540])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..080769f86717bf9ba3f1a0426cbbf4a7b201a07d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_tensor([3540, 3540])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tense, low-angle, handheld, Steadicam tracking shot closely follows a young woman as she dashes with agility and power through the ruins of an ancient castle, densely covered in vines. Her dark brown hair, worn in a high ponytail, a few sweat-soaked strands plastered to her face in a focused, anxious state. Her lips are pursed, and her eyes scan the obstacle-filled path ahead with a sharp gaze. She wears a fitted charcoal gray overalls, sleeves casually rolled to the elbows, and a pair of worn brown leather hiking boots. Each step is precise and swift over the scattered rubble and broken stone pillars. Golden afternoon sunlight filters through the crumbling Gothic arches and gaping stone walls overhead, slicing through the air thick with ancient dust, creating distinct, Tyndall-effect beams that oscillate with the camera's movement, casting shifting, dappled shadows across the ancient stone pavement. The entire image is imbued with a cinematic quality, with sharp focus and a subtle sense of film grain, perfectly capturing the dynamic beauty and suspenseful tension of this thrilling chase.", + "seed": 13354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_3540_seed_13354_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027059067133814098, + "vlm_logprob": -0.027059067133814098, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.027059067133814098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016152240335941315 + ], + "opposite_logprob_tokens": [ + -4.266152381896973 + ], + "logprob_reward": -0.016152240335941315, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016152240335941315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10779925435781479, + 0.05447956547141075, + 0.03514911234378815, + 0.1654360294342041, + 0.05533554404973984, + 0.08192433416843414, + 0.09733042120933533, + 0.06723862886428833, + 0.1069330945611, + 0.22837400436401367 + ], + "credit_map": [ + 0.014287074096500874, + 0.00988153275102377, + 0.017754897475242615, + 0.015892112627625465, + 0.009499805979430676, + 0.01106597576290369, + 0.011765619739890099, + 0.017652232199907303, + 0.0076703657396137714, + 0.003965691663324833, + 0.014488072134554386, + 0.0074930861592292786, + 0.0071450937539339066, + 0.004117534030228853, + 0.004799674265086651, + 0.004800048656761646, + 0.004852588288486004, + 0.003683132817968726, + 0.0055723292753100395, + 0.003937105648219585, + 0.00461329473182559, + 0.003107525408267975, + 0.003049984574317932, + 0.006333151366561651, + 0.030278855934739113, + 0.01248850580304861, + 0.020735738798975945, + 0.019197536632418633, + 0.040462519973516464, + 0.01151072047650814, + 0.019959421828389168, + 0.010802733711898327, + 0.004121563863009214, + 0.005493583157658577, + 0.008159968070685863, + 0.006040402688086033, + 0.01038979459553957, + 0.005206827539950609, + 0.011742007918655872, + 0.00418139947578311, + 0.00684773363173008, + 0.014183973893523216, + 0.008306486532092094, + 0.015509089455008507, + 0.01398899033665657, + 0.00547050591558218, + 0.012354887090623379, + 0.005262657534331083, + 0.010140261612832546, + 0.00876416265964508, + 0.013997216708958149, + 0.015623033046722412, + 0.01598930172622204, + 0.006963424850255251, + 0.013056729920208454, + 0.012796295806765556, + 0.007725827861577272, + 0.011766605079174042, + 0.010687526315450668, + 0.01063875388354063, + 0.008430149406194687, + 0.004696037154644728, + 0.007869916036725044, + 0.005423808936029673, + 0.0075796558521687984, + 0.009673379361629486, + 0.014593790285289288, + 0.014628234319388866, + 0.018351972103118896, + 0.0150604872033, + 0.01536937803030014, + 0.01167619600892067, + 0.034064881503582, + 0.034712053835392, + 0.03598199039697647, + 0.02342928946018219, + 0.037426844239234924, + 0.01925905980169773, + 0.02276056818664074, + 0.020739324390888214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.016152240335941315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348866522312164 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348866522312164, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348866522312164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0942743718624115, + 0.021560203284025192, + 0.02361641265451908, + 0.12216584384441376, + 0.02229364402592182, + 0.06952884048223495, + 0.039629630744457245, + 0.026258347555994987, + 0.04341597110033035, + 0.5372567176818848 + ], + "credit_map": [ + 0.008602726273238659, + 0.006229943595826626, + 0.0234657134860754, + 0.008493679575622082, + 0.007599376142024994, + 0.006771527696400881, + 0.007933014072477818, + 0.025178387761116028, + 0.002632718998938799, + 0.0015075027476996183, + 0.003035148372873664, + 0.002233120147138834, + 0.003346238052472472, + 0.00243879621848464, + 0.0034743240103125572, + 0.0028923542704433203, + 0.003403228707611561, + 0.0017459727823734283, + 0.004048031754791737, + 0.002348753158003092, + 0.0030560879968106747, + 0.0036347401328384876, + 0.002370321424677968, + 0.003009276231750846, + 0.022449567914009094, + 0.009673415683209896, + 0.015428655780851841, + 0.013943260535597801, + 0.029412975534796715, + 0.00914595928043127, + 0.014271372929215431, + 0.007840634323656559, + 0.002146493410691619, + 0.002134860260412097, + 0.0025206408463418484, + 0.004161953926086426, + 0.0029451665468513966, + 0.001990992808714509, + 0.0036244529765099287, + 0.002769084880128503, + 0.007818133570253849, + 0.00976435374468565, + 0.009180305525660515, + 0.021502073854207993, + 0.007851097732782364, + 0.0030527706257998943, + 0.006360260769724846, + 0.0039998386055231094, + 0.0033354703336954117, + 0.004362845327705145, + 0.007723715156316757, + 0.005707136821001768, + 0.004424439277499914, + 0.0038406806997954845, + 0.006304789334535599, + 0.0039305537939071655, + 0.003426137613132596, + 0.003593409899622202, + 0.004050161689519882, + 0.0029020304791629314, + 0.0035103813279420137, + 0.0033757665660232306, + 0.0030276316683739424, + 0.0023728273808956146, + 0.005682794842869043, + 0.004218944348394871, + 0.006127900443971157, + 0.007044878322631121, + 0.004343738779425621, + 0.005730829667299986, + 0.006374947261065245, + 0.0038919332437217236, + 0.12807975709438324, + 0.06520675122737885, + 0.10684284567832947, + 0.046048495918512344, + 0.04231921210885048, + 0.05778651311993599, + 0.0422721765935421, + 0.048700980842113495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.026348866522312164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605992205440998 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.01605992205440998, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605992205440998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22418780624866486, + 0.17637889087200165, + 0.044379089027643204, + 0.17617540061473846, + 0.018086766824126244, + 0.06302843987941742, + 0.024210156872868538, + 0.06640076637268066, + 0.02420903369784355, + 0.1829436868429184 + ], + "credit_map": [ + 0.021336980164051056, + 0.014433764852583408, + 0.05237714946269989, + 0.019937502220273018, + 0.026192931458353996, + 0.017102887853980064, + 0.014683762565255165, + 0.058122795075178146, + 0.020933425053954124, + 0.011096195317804813, + 0.017579087987542152, + 0.016545863822102547, + 0.044833432883024216, + 0.017588496208190918, + 0.028740357607603073, + 0.019062016159296036, + 0.007104883901774883, + 0.004555497784167528, + 0.007016077172011137, + 0.005115927197039127, + 0.005407856311649084, + 0.006473209243267775, + 0.004455043468624353, + 0.004250589292496443, + 0.04301559180021286, + 0.015208031982183456, + 0.020771371200680733, + 0.024569518864154816, + 0.025751881301403046, + 0.01487888116389513, + 0.019884346053004265, + 0.01209576427936554, + 0.00213451124727726, + 0.002049395116046071, + 0.0022845643106848, + 0.002620530780404806, + 0.002524926792830229, + 0.0016116639599204063, + 0.0027536191046237946, + 0.0021075555123388767, + 0.00838757585734129, + 0.014247466810047626, + 0.006677879486232996, + 0.007220273371785879, + 0.009976048022508621, + 0.006173437461256981, + 0.005777124781161547, + 0.004568629432469606, + 0.002678028540685773, + 0.003877002978697419, + 0.0037597662303596735, + 0.0034163719974458218, + 0.002858288586139679, + 0.002768218982964754, + 0.0026678743306547403, + 0.0021846024319529533, + 0.007534475065767765, + 0.010102623142302036, + 0.009304435923695564, + 0.010762959718704224, + 0.0056332736276090145, + 0.00896122120320797, + 0.005613244138658047, + 0.008488528430461884, + 0.0033217512536793947, + 0.002334634540602565, + 0.0029334770515561104, + 0.003483779728412628, + 0.0029218175914138556, + 0.0037781046703457832, + 0.0031882827170193195, + 0.0022471838165074587, + 0.04202908277511597, + 0.022533705458045006, + 0.03552405536174774, + 0.016612576320767403, + 0.015506330877542496, + 0.01900292932987213, + 0.014206869527697563, + 0.017528126016259193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01605992205440998 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42879459261894226, + 0.042798466980457306, + 0.02935105934739113, + 0.06465652585029602, + 0.0735522210597992, + 0.04764031991362572, + 0.05473830923438072, + 0.03984958678483963, + 0.17491964995861053, + 0.0436992347240448 + ], + "credit_map": [ + 0.0387837179005146, + 0.025888213887810707, + 0.12330986559391022, + 0.029538648203015327, + 0.01763751544058323, + 0.025564847514033318, + 0.017980579286813736, + 0.15009121596813202, + 0.0049706571735441685, + 0.004848402924835682, + 0.008404113352298737, + 0.007050590123981237, + 0.005498990416526794, + 0.0039733476005494595, + 0.0034321185667067766, + 0.004620250780135393, + 0.005638183560222387, + 0.004325584974139929, + 0.0034330396447330713, + 0.004033876117318869, + 0.0030650279950350523, + 0.0024473124649375677, + 0.0032177448738366365, + 0.003190288320183754, + 0.01047301571816206, + 0.006729567423462868, + 0.007746662013232708, + 0.0066579412668943405, + 0.01575149968266487, + 0.00476303044706583, + 0.007747528608888388, + 0.0047872792929410934, + 0.01228036917746067, + 0.013182424940168858, + 0.008703228086233139, + 0.01134260930120945, + 0.00832939613610506, + 0.005310938693583012, + 0.008683712221682072, + 0.005719544366002083, + 0.0051507302559912205, + 0.009440752677619457, + 0.005762769374996424, + 0.012452177703380585, + 0.004342706874012947, + 0.0028836261481046677, + 0.0044899582862854, + 0.0031176006887108088, + 0.0064827147871255875, + 0.0073352959007024765, + 0.007753679528832436, + 0.009299657307565212, + 0.008642454631626606, + 0.005419508088380098, + 0.004890507087111473, + 0.004914492834359407, + 0.005218719597905874, + 0.007967631332576275, + 0.0060068960301578045, + 0.005749393720179796, + 0.0032388565596193075, + 0.0034491971600800753, + 0.0036942369770258665, + 0.004524655640125275, + 0.032282594591379166, + 0.017861269414424896, + 0.02934246137738228, + 0.020743580535054207, + 0.011104004457592964, + 0.025835799053311348, + 0.018725819885730743, + 0.01902410387992859, + 0.007938330993056297, + 0.00677231652662158, + 0.006564850453287363, + 0.004374236799776554, + 0.005292237736284733, + 0.005031340289860964, + 0.003853035159409046, + 0.003872890956699848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999847292900085 + ], + "opposite_logprob_tokens": [ + -2.6949985027313232 + ], + "logprob_reward": -0.06999847292900085, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999847292900085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10405208170413971, + 0.054415468126535416, + 0.04855865240097046, + 0.18317405879497528, + 0.06932814419269562, + 0.07767577469348907, + 0.23155300319194794, + 0.05253877490758896, + 0.07583605498075485, + 0.1028679609298706 + ], + "credit_map": [ + 0.01270238682627678, + 0.007732919417321682, + 0.013694286346435547, + 0.017092375084757805, + 0.015020079910755157, + 0.011654081754386425, + 0.011495528742671013, + 0.01466042548418045, + 0.006487497594207525, + 0.004179708193987608, + 0.00735994428396225, + 0.006800511851906776, + 0.011275780387222767, + 0.00552734499797225, + 0.006521030329167843, + 0.0062636518850922585, + 0.008744284510612488, + 0.005225616507232189, + 0.007517905905842781, + 0.005914650857448578, + 0.005267379805445671, + 0.0065611908212304115, + 0.00476741511374712, + 0.004560206551104784, + 0.033102720975875854, + 0.019874026998877525, + 0.022480301558971405, + 0.020645756274461746, + 0.04056169092655182, + 0.014289772137999535, + 0.020507819950580597, + 0.0117119699716568, + 0.008696592412889004, + 0.008358524180948734, + 0.007144799921661615, + 0.01213111262768507, + 0.011148548685014248, + 0.006386167835444212, + 0.009489000774919987, + 0.00597339728847146, + 0.008588908240199089, + 0.012013797648251057, + 0.008378585800528526, + 0.011447671800851822, + 0.014752608723938465, + 0.005284342914819717, + 0.011549713090062141, + 0.005660150200128555, + 0.020603002980351448, + 0.02397928386926651, + 0.03746648505330086, + 0.03956369310617447, + 0.03198268637061119, + 0.01848137378692627, + 0.03873597830533981, + 0.020740509033203125, + 0.004780821036547422, + 0.00725314998999238, + 0.007622408214956522, + 0.009262990206480026, + 0.004916542675346136, + 0.004750853404402733, + 0.007140686269849539, + 0.006811325438320637, + 0.011891152709722519, + 0.00765302125364542, + 0.01071097794920206, + 0.009939098730683327, + 0.005220750812441111, + 0.011591934598982334, + 0.00994538888335228, + 0.008883731439709663, + 0.02154402993619442, + 0.014318989589810371, + 0.014807895757257938, + 0.0088478559628129, + 0.011064759455621243, + 0.012565533630549908, + 0.007828310132026672, + 0.011890586465597153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.06999847292900085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, low-angle, handheld, Steadicam tracking shot closely follows a young woman as she dashes with agility and power through the ruins of an ancient castle, densely covered in vines. Her dark brown hair, worn in a high ponytail, a few sweat-soaked strands plastered to her face in a focused, anxious state. Her lips are pursed, and her eyes scan the obstacle-filled path ahead with a sharp gaze. She wears a fitted charcoal gray overalls, sleeves casually rolled to the elbows, and a pair of worn brown leather hiking boots. Each step is precise and swift over the scattered rubble and broken stone pillars. Golden afternoon sunlight filters through the crumbling Gothic arches and gaping stone walls overhead, slicing through the air thick with ancient dust, creating distinct, Tyndall-effect beams that oscillate with the camera's movement, casting shifting, dappled shadows across the ancient stone pavement. The entire image is imbued with a cinematic quality, with sharp focus and a subtle sense of film grain, perfectly capturing the dynamic beauty and suspenseful tension of this thrilling chase.", + "seed": 13355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_3540_seed_13355_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017577502876520157, + "vlm_logprob": -0.017577502876520157, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.017577502876520157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01426651980727911 + ], + "opposite_logprob_tokens": [ + -4.389266490936279 + ], + "logprob_reward": -0.01426651980727911, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01426651980727911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2535799741744995, + 0.037184152752161026, + 0.14523661136627197, + 0.07581660896539688, + 0.04478960856795311, + 0.09848066419363022, + 0.07451777160167694, + 0.0999414324760437, + 0.08664355427026749, + 0.08380956947803497 + ], + "credit_map": [ + 0.01924438402056694, + 0.00966316182166338, + 0.02572842687368393, + 0.013530733063817024, + 0.023358026519417763, + 0.07393775880336761, + 0.018903454765677452, + 0.06921402364969254, + 0.003927462734282017, + 0.0046067992225289345, + 0.004166710190474987, + 0.0041267601773142815, + 0.004684934858232737, + 0.00518154026940465, + 0.006313729099929333, + 0.004176218993961811, + 0.011164369992911816, + 0.010341488756239414, + 0.041510939598083496, + 0.017074674367904663, + 0.004729790613055229, + 0.008980343118309975, + 0.007301541976630688, + 0.04413347691297531, + 0.005538296420127153, + 0.007479551248252392, + 0.009205714799463749, + 0.010607771575450897, + 0.010977586731314659, + 0.009340084157884121, + 0.013209040276706219, + 0.009458569809794426, + 0.0046933856792747974, + 0.0031280862167477608, + 0.005858022719621658, + 0.007875487208366394, + 0.00506978202611208, + 0.006274484097957611, + 0.004635846707969904, + 0.0072545199654996395, + 0.01083570159971714, + 0.007523694541305304, + 0.013159834779798985, + 0.018534917384386063, + 0.012196844443678856, + 0.010445072315633297, + 0.013223160058259964, + 0.012561448849737644, + 0.007184513844549656, + 0.011260343715548515, + 0.008481166325509548, + 0.010516034439206123, + 0.009145276620984077, + 0.007209595758467913, + 0.011433526873588562, + 0.009287318214774132, + 0.01279755774885416, + 0.009214967489242554, + 0.013952995650470257, + 0.0139310909435153, + 0.01446689572185278, + 0.011552796699106693, + 0.012175883166491985, + 0.011849251575767994, + 0.007473033852875233, + 0.014096747152507305, + 0.017578493803739548, + 0.011737012304365635, + 0.009035862050950527, + 0.0066863251850008965, + 0.01055222563445568, + 0.009483857080340385, + 0.0051249307580292225, + 0.022018587216734886, + 0.022188575938344002, + 0.00899833906441927, + 0.006648470181971788, + 0.0052695488557219505, + 0.0069701713509857655, + 0.006590959150344133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.01426651980727911 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34551218152046204, + 0.08327697962522507, + 0.04714975506067276, + 0.08050080388784409, + 0.04498480632901192, + 0.05258649215102196, + 0.07854140549898148, + 0.11428903043270111, + 0.06616032868623734, + 0.08699826896190643 + ], + "credit_map": [ + 0.02375207282602787, + 0.014875689521431923, + 0.1112012267112732, + 0.023324334993958473, + 0.024516481906175613, + 0.030706113204360008, + 0.021900877356529236, + 0.09523536264896393, + 0.011015239171683788, + 0.008785313926637173, + 0.0072826966643333435, + 0.006988824810832739, + 0.01559423003345728, + 0.010964379645884037, + 0.012963613495230675, + 0.009682677686214447, + 0.003639885690063238, + 0.005269114393740892, + 0.008086412213742733, + 0.006498127244412899, + 0.0043905917555093765, + 0.003514875890687108, + 0.010268401354551315, + 0.005482336971908808, + 0.0078507661819458, + 0.01045731920748949, + 0.0103215416893363, + 0.014104383997619152, + 0.01104673370718956, + 0.00893616396933794, + 0.011814567260444164, + 0.005969315301626921, + 0.004113623406738043, + 0.0029901210218667984, + 0.004581032320857048, + 0.006175218615680933, + 0.010115020908415318, + 0.004263672046363354, + 0.0061985584907233715, + 0.006547550670802593, + 0.007009638473391533, + 0.009267447516322136, + 0.005862637422978878, + 0.007223009597510099, + 0.005175033118575811, + 0.005355076864361763, + 0.0063835568726062775, + 0.006310086231678724, + 0.00774276303127408, + 0.010697251185774803, + 0.007984803058207035, + 0.010610257275402546, + 0.010346993803977966, + 0.006274606566876173, + 0.016327595338225365, + 0.008557128719985485, + 0.01572330854833126, + 0.010016717948019505, + 0.02041623741388321, + 0.01571650616824627, + 0.015252604149281979, + 0.007490510120987892, + 0.017355388030409813, + 0.012317738495767117, + 0.009122096933424473, + 0.006254904437810183, + 0.015229451470077038, + 0.012316488660871983, + 0.005000845063477755, + 0.005631419364362955, + 0.007052397355437279, + 0.005552712827920914, + 0.006732983980327845, + 0.018375039100646973, + 0.022442590445280075, + 0.00890661496669054, + 0.008224369958043098, + 0.007808004971593618, + 0.008118605241179466, + 0.006390056107193232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631303906440735 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631303906440735, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631303906440735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43614479899406433, + 0.051206644624471664, + 0.12780305743217468, + 0.09952805191278458, + 0.04065806046128273, + 0.03666175156831741, + 0.03647719323635101, + 0.030254220589995384, + 0.09663520753383636, + 0.0446309857070446 + ], + "credit_map": [ + 0.025558223947882652, + 0.015113363042473793, + 0.1256994754076004, + 0.029469281435012817, + 0.02249777317047119, + 0.06319992244243622, + 0.0204820167273283, + 0.134124755859375, + 0.007002444937825203, + 0.0077371858060359955, + 0.005100311245769262, + 0.0060315621085464954, + 0.005445757415145636, + 0.010007808916270733, + 0.004553355276584625, + 0.005328222643584013, + 0.015198631212115288, + 0.012228150852024555, + 0.023191284388303757, + 0.016026655212044716, + 0.008735843934118748, + 0.01187934074550867, + 0.010179093107581139, + 0.03036407195031643, + 0.010859244503080845, + 0.01603027619421482, + 0.015246989205479622, + 0.015509226359426975, + 0.010280722752213478, + 0.012655537575483322, + 0.010665908455848694, + 0.00828014686703682, + 0.0039869449101388454, + 0.002553836675360799, + 0.0043787541799247265, + 0.006998784840106964, + 0.006497714668512344, + 0.0037688857410103083, + 0.0048939879052340984, + 0.007579151540994644, + 0.005698690190911293, + 0.007342074532061815, + 0.005044121295213699, + 0.007460600230842829, + 0.0029258746653795242, + 0.0019762334413826466, + 0.0032789858523756266, + 0.002935170428827405, + 0.0035292853135615587, + 0.006862153764814138, + 0.004535147920250893, + 0.0050626713782548904, + 0.004960713908076286, + 0.002362277824431658, + 0.004608053248375654, + 0.004556891042739153, + 0.0031778179109096527, + 0.0050307936035096645, + 0.0030087216291576624, + 0.006080219987779856, + 0.004499590955674648, + 0.00268914015032351, + 0.0035452493466436863, + 0.002222687704488635, + 0.010606943629682064, + 0.015731751918792725, + 0.024841947481036186, + 0.016984257847070694, + 0.005995857063680887, + 0.007934228517115116, + 0.008845550008118153, + 0.005694663152098656, + 0.002953204093500972, + 0.0106468191370368, + 0.012019387446343899, + 0.0058262706734240055, + 0.004346103873103857, + 0.0023745503276586533, + 0.004169846884906292, + 0.0022948035039007664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008631303906440735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2206949144601822, + 0.11772388964891434, + 0.0592796728014946, + 0.07181353867053986, + 0.17889821529388428, + 0.15893404185771942, + 0.0343589261174202, + 0.056122127920389175, + 0.0655493512749672, + 0.03662532940506935 + ], + "credit_map": [ + 0.018181150779128075, + 0.012159952893853188, + 0.07646578550338745, + 0.02059571072459221, + 0.01855529099702835, + 0.021414797753095627, + 0.014334612525999546, + 0.03898760676383972, + 0.01634928397834301, + 0.02172745205461979, + 0.012991497293114662, + 0.014182991348206997, + 0.00868943426758051, + 0.01676742359995842, + 0.01093248464167118, + 0.01608332060277462, + 0.004571088124066591, + 0.005694628693163395, + 0.008259202353656292, + 0.009366188198328018, + 0.00452807592228055, + 0.005817091558128595, + 0.009556280449032784, + 0.011487115174531937, + 0.00846643466502428, + 0.010514527559280396, + 0.009999853558838367, + 0.010088718496263027, + 0.008961329236626625, + 0.007831909693777561, + 0.007932404987514019, + 0.00801835861057043, + 0.019245512783527374, + 0.011778300628066063, + 0.019013991579413414, + 0.0388190932571888, + 0.03065582364797592, + 0.014280060306191444, + 0.021364035084843636, + 0.023741405457258224, + 0.034565042704343796, + 0.022744053974747658, + 0.026075612753629684, + 0.03317758068442345, + 0.008421139791607857, + 0.009591372683644295, + 0.012839416041970253, + 0.011519821360707283, + 0.0023229955695569515, + 0.0037430597003549337, + 0.005466068629175425, + 0.007326338440179825, + 0.004725046921521425, + 0.0021263770759105682, + 0.005374434869736433, + 0.0032746042124927044, + 0.008146620355546474, + 0.007592692039906979, + 0.01045713946223259, + 0.008580231107771397, + 0.0061851805076003075, + 0.004014953970909119, + 0.0059983697719872, + 0.005146937444806099, + 0.007380127441138029, + 0.012519314885139465, + 0.008479718118906021, + 0.012724095955491066, + 0.005255349446088076, + 0.005541674327105284, + 0.006673992145806551, + 0.006975078955292702, + 0.0034653954207897186, + 0.003965322859585285, + 0.004357564728707075, + 0.006614710204303265, + 0.004997059237211943, + 0.005136969033628702, + 0.004792188759893179, + 0.0032961179967969656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0036151548847556114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808215260505676 + ], + "opposite_logprob_tokens": [ + -3.2880821228027344 + ], + "logprob_reward": -0.03808215260505676, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808215260505676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4987662136554718, + 0.05316965654492378, + 0.06288167089223862, + 0.03498517721891403, + 0.10170070081949234, + 0.06006850674748421, + 0.042851854115724564, + 0.031254593282938004, + 0.08755383640527725, + 0.026767781004309654 + ], + "credit_map": [ + 0.032027170062065125, + 0.013871237635612488, + 0.1855187565088272, + 0.036379601806402206, + 0.028713561594486237, + 0.03608976677060127, + 0.03017244301736355, + 0.13599371910095215, + 0.006195512134581804, + 0.004956827964633703, + 0.004558534827083349, + 0.0048540313728153706, + 0.01019737683236599, + 0.007402059622108936, + 0.008173891343176365, + 0.006831429433077574, + 0.005159619729965925, + 0.008093391545116901, + 0.017748797312378883, + 0.008064871653914452, + 0.003037268528714776, + 0.0034378955606371164, + 0.0047858962789177895, + 0.012553934007883072, + 0.0035487357527017593, + 0.004554238636046648, + 0.0053405119106173515, + 0.004904449917376041, + 0.004291624762117863, + 0.0034488264936953783, + 0.005334764253348112, + 0.003562026657164097, + 0.011342418380081654, + 0.007918916642665863, + 0.014279179275035858, + 0.017337741330266, + 0.010681952349841595, + 0.008949385955929756, + 0.011456327512860298, + 0.01973479613661766, + 0.008431671187281609, + 0.008689258247613907, + 0.008611474186182022, + 0.010160380974411964, + 0.005829567555338144, + 0.005538496654480696, + 0.006508646998554468, + 0.006299019791185856, + 0.0027776407077908516, + 0.00487983925268054, + 0.006096117198467255, + 0.008834960870444775, + 0.006874813698232174, + 0.00352070783264935, + 0.005487208720296621, + 0.004380570724606514, + 0.00421726331114769, + 0.004756398499011993, + 0.005236965604126453, + 0.004032375756651163, + 0.003417238825932145, + 0.0019333693198859692, + 0.00425906665623188, + 0.0034019220620393753, + 0.006806439720094204, + 0.020719675347208977, + 0.022394323721528053, + 0.009955838322639465, + 0.00603820662945509, + 0.0054047973826527596, + 0.009304368868470192, + 0.006930192932486534, + 0.0029055748600512743, + 0.0038664257153868675, + 0.0041882432997226715, + 0.0033791796304285526, + 0.003137497929856181, + 0.003754471894353628, + 0.0029688419308513403, + 0.0025675473734736443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03808215260505676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, low-angle, handheld, Steadicam tracking shot closely follows a young woman as she dashes with agility and power through the ruins of an ancient castle, densely covered in vines. Her dark brown hair, worn in a high ponytail, a few sweat-soaked strands plastered to her face in a focused, anxious state. Her lips are pursed, and her eyes scan the obstacle-filled path ahead with a sharp gaze. She wears a fitted charcoal gray overalls, sleeves casually rolled to the elbows, and a pair of worn brown leather hiking boots. Each step is precise and swift over the scattered rubble and broken stone pillars. Golden afternoon sunlight filters through the crumbling Gothic arches and gaping stone walls overhead, slicing through the air thick with ancient dust, creating distinct, Tyndall-effect beams that oscillate with the camera's movement, casting shifting, dappled shadows across the ancient stone pavement. The entire image is imbued with a cinematic quality, with sharp focus and a subtle sense of film grain, perfectly capturing the dynamic beauty and suspenseful tension of this thrilling chase.", + "seed": 13356, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_3540_seed_13356_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018780254852026702, + "vlm_logprob": -0.018780254852026702, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.018780254852026702, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01924649439752102 + ], + "opposite_logprob_tokens": [ + -4.019246578216553 + ], + "logprob_reward": -0.01924649439752102, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01924649439752102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06775108724832535, + 0.03633151575922966, + 0.03152167424559593, + 0.02823995053768158, + 0.05944610387086868, + 0.07019852101802826, + 0.09816552698612213, + 0.12771382927894592, + 0.05998227372765541, + 0.42064955830574036 + ], + "credit_map": [ + 0.0102050406858325, + 0.011426540091633797, + 0.008663421496748924, + 0.007359732873737812, + 0.007534800563007593, + 0.0086184898391366, + 0.007888739928603172, + 0.0060543217696249485, + 0.003035149769857526, + 0.002919405698776245, + 0.008829288184642792, + 0.003918591421097517, + 0.002968121087178588, + 0.0044036805629730225, + 0.0026124597061425447, + 0.007644818164408207, + 0.0033798490185290575, + 0.004103763960301876, + 0.003102621529251337, + 0.0027554621919989586, + 0.006524498574435711, + 0.005241721402853727, + 0.004022557288408279, + 0.0023911993484944105, + 0.0030464655719697475, + 0.002721390686929226, + 0.0033639727625995874, + 0.0026330542750656605, + 0.00528310053050518, + 0.003705904819071293, + 0.005262705497443676, + 0.002223354997113347, + 0.005920175462961197, + 0.005986941047012806, + 0.006458345800638199, + 0.005403722170740366, + 0.010490048676729202, + 0.008947471156716347, + 0.00927003100514412, + 0.006969365291297436, + 0.005953919608145952, + 0.00789722055196762, + 0.008917149156332016, + 0.011148982681334019, + 0.01280803233385086, + 0.004580638371407986, + 0.012127423658967018, + 0.0067651597782969475, + 0.011845482513308525, + 0.008172119967639446, + 0.013216954655945301, + 0.017105458304286003, + 0.00954389851540327, + 0.006044068839401007, + 0.019521033391356468, + 0.012716507539153099, + 0.016416968777775764, + 0.010597956366837025, + 0.023381125181913376, + 0.019077787175774574, + 0.013559741899371147, + 0.007766012568026781, + 0.02074844017624855, + 0.016165796667337418, + 0.004234026186168194, + 0.009215989150106907, + 0.007365557365119457, + 0.010439150035381317, + 0.007442563306540251, + 0.005429113749414682, + 0.008877001702785492, + 0.0069788675755262375, + 0.038428500294685364, + 0.05974982678890228, + 0.03891562297940254, + 0.047939591109752655, + 0.039902109652757645, + 0.07548508793115616, + 0.04328131675720215, + 0.07694748789072037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.01924649439752102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819327473640442 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819327473640442, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819327473640442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12621966004371643, + 0.14206841588020325, + 0.1572423130273819, + 0.06943941116333008, + 0.052375949919223785, + 0.06201881542801857, + 0.14977987110614777, + 0.09034966677427292, + 0.0475221611559391, + 0.1029837355017662 + ], + "credit_map": [ + 0.016096165403723717, + 0.013898342847824097, + 0.01921107806265354, + 0.014506925828754902, + 0.012740482576191425, + 0.02327977679669857, + 0.012762063182890415, + 0.013724819757044315, + 0.009544233791530132, + 0.01933867856860161, + 0.017323506996035576, + 0.014008572325110435, + 0.02900555171072483, + 0.02230660803616047, + 0.014435858465731144, + 0.0161054078489542, + 0.015195195563137531, + 0.02160419523715973, + 0.019683020189404488, + 0.012959365732967854, + 0.022182270884513855, + 0.036678146570920944, + 0.01677720621228218, + 0.012162910774350166, + 0.005264499224722385, + 0.0076794675551354885, + 0.008772233501076698, + 0.006184043828397989, + 0.01484172698110342, + 0.007499415427446365, + 0.014182565733790398, + 0.005015457049012184, + 0.004919308237731457, + 0.006508176214993, + 0.00567645812407136, + 0.0038314522244036198, + 0.009799827821552753, + 0.009742814116179943, + 0.00563490716740489, + 0.006263006944209337, + 0.007051948457956314, + 0.009587723761796951, + 0.0073170894756913185, + 0.008886093273758888, + 0.008733092807233334, + 0.004758263472467661, + 0.009832130745053291, + 0.005852475296705961, + 0.01822289638221264, + 0.01715889759361744, + 0.01977926678955555, + 0.02444586344063282, + 0.013471730053424835, + 0.00852668285369873, + 0.02893691323697567, + 0.019237613305449486, + 0.009978792630136013, + 0.006234950851649046, + 0.013979759067296982, + 0.014932137914001942, + 0.01295888889580965, + 0.004120682366192341, + 0.02048240415751934, + 0.007662054151296616, + 0.003083079354837537, + 0.008476423099637032, + 0.0046903593465685844, + 0.010023020207881927, + 0.004781247116625309, + 0.004137622658163309, + 0.006542746443301439, + 0.005787665955722332, + 0.006856172811239958, + 0.01374804601073265, + 0.009475004859268665, + 0.01275299396365881, + 0.012195277027785778, + 0.018658237531781197, + 0.014131140895187855, + 0.01516686100512743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.01819327473640442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23642992973327637, + 0.34424206614494324, + 0.10115940868854523, + 0.0569012276828289, + 0.03366803005337715, + 0.03204616904258728, + 0.03873513266444206, + 0.048470426350831985, + 0.0651480183005333, + 0.04319960996508598 + ], + "credit_map": [ + 0.03523479029536247, + 0.033364441245794296, + 0.029900791123509407, + 0.027016256004571915, + 0.02009102888405323, + 0.03218621015548706, + 0.025055687874555588, + 0.0335807166993618, + 0.0302216038107872, + 0.04366186633706093, + 0.09466160088777542, + 0.028904464095830917, + 0.03519602492451668, + 0.060031551867723465, + 0.022368330508470535, + 0.029196593910455704, + 0.01769341342151165, + 0.016945553943514824, + 0.011820490472018719, + 0.010157723911106586, + 0.01083674468100071, + 0.017080070450901985, + 0.007974597625434399, + 0.00865081138908863, + 0.009547914378345013, + 0.007994597777724266, + 0.007852671667933464, + 0.007552223280072212, + 0.005638376343995333, + 0.007079171482473612, + 0.005515919532626867, + 0.005720353219658136, + 0.003304491750895977, + 0.004710517358034849, + 0.004462671000510454, + 0.0039201946929097176, + 0.003191102296113968, + 0.005485107656568289, + 0.002889591734856367, + 0.00570435356348753, + 0.006048883311450481, + 0.0049210297875106335, + 0.0041107358410954475, + 0.004319163970649242, + 0.003293284447863698, + 0.0017257868312299252, + 0.004652279429137707, + 0.0029750047251582146, + 0.0034996371250599623, + 0.003168924944475293, + 0.005003007128834724, + 0.00585694657638669, + 0.0038425023667514324, + 0.0032150938641279936, + 0.009488258510828018, + 0.0046607619151473045, + 0.005982568953186274, + 0.004701383411884308, + 0.0062903184443712234, + 0.006961436476558447, + 0.006284754723310471, + 0.004357422236353159, + 0.007068234495818615, + 0.006824307609349489, + 0.005516601260751486, + 0.005191918928176165, + 0.012719529680907726, + 0.007163007277995348, + 0.008138669654726982, + 0.011569341644644737, + 0.007437000051140785, + 0.0074119544588029385, + 0.0035268114879727364, + 0.005997929256409407, + 0.003320941934362054, + 0.006573044694960117, + 0.005119465757161379, + 0.0064478605054318905, + 0.007134387269616127, + 0.005079166032373905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004633048549294472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031916177831590176 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031916177831590176, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031916177831590176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19126805663108826, + 0.30127206444740295, + 0.22083909809589386, + 0.06739389896392822, + 0.03194719925522804, + 0.055036090314388275, + 0.035209886729717255, + 0.03221855312585831, + 0.02298746444284916, + 0.04182768613100052 + ], + "credit_map": [ + 0.021707070991396904, + 0.026393700391054153, + 0.018986549228429794, + 0.02740812860429287, + 0.02380300872027874, + 0.03729662299156189, + 0.018324147909879684, + 0.017348816618323326, + 0.037015803158283234, + 0.0431177094578743, + 0.09224580973386765, + 0.02518320642411709, + 0.02891351841390133, + 0.04628560692071915, + 0.011617681942880154, + 0.01689271815121174, + 0.0268949456512928, + 0.0501328781247139, + 0.023083290085196495, + 0.01583845727145672, + 0.033512696623802185, + 0.03744501993060112, + 0.013522762805223465, + 0.020409023389220238, + 0.010010145604610443, + 0.011871772818267345, + 0.009674539789557457, + 0.008218272589147091, + 0.005923268850892782, + 0.008531632833182812, + 0.006216580513864756, + 0.006947685964405537, + 0.0037399213761091232, + 0.0045987931080162525, + 0.004656018689274788, + 0.003803274594247341, + 0.0027877273969352245, + 0.004540383815765381, + 0.0028315542731434107, + 0.004989526234567165, + 0.008429693058133125, + 0.005368840880692005, + 0.007539781276136637, + 0.010008426383137703, + 0.004421568941324949, + 0.00341969751752913, + 0.006945962086319923, + 0.008902113884687424, + 0.00444608461111784, + 0.003601596923545003, + 0.004630636423826218, + 0.007417806424200535, + 0.004679958801716566, + 0.002326592104509473, + 0.004386383108794689, + 0.0037208322901278734, + 0.0035428395494818687, + 0.0029221943113952875, + 0.004189374856650829, + 0.004709383938461542, + 0.004472317639738321, + 0.0030171265825629234, + 0.004861405119299889, + 0.0045039113610982895, + 0.001332800486125052, + 0.003892238950356841, + 0.0024031223729252815, + 0.004123148508369923, + 0.0025156366173177958, + 0.002037579193711281, + 0.003971908241510391, + 0.0027110297232866287, + 0.003934613429009914, + 0.005972304381430149, + 0.0037419823929667473, + 0.0060430062003433704, + 0.004461369477212429, + 0.005489014089107513, + 0.007102676201611757, + 0.005082719959318638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0031916177831590176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863683879375458 + ], + "opposite_logprob_tokens": [ + -3.0486369132995605 + ], + "logprob_reward": -0.04863683879375458, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863683879375458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2114911526441574, + 0.10505946725606918, + 0.0993977040052414, + 0.1218169704079628, + 0.06946483254432678, + 0.0656980648636818, + 0.1468329131603241, + 0.06421416252851486, + 0.052887238562107086, + 0.06313744932413101 + ], + "credit_map": [ + 0.030540665611624718, + 0.030107738450169563, + 0.02389897033572197, + 0.023676270619034767, + 0.026129761710762978, + 0.03769276291131973, + 0.017093073576688766, + 0.022351926192641258, + 0.006556544452905655, + 0.011112851090729237, + 0.031099261716008186, + 0.007574338931590319, + 0.016122162342071533, + 0.01634819246828556, + 0.006489092949777842, + 0.009757021442055702, + 0.010309379547834396, + 0.013049610890448093, + 0.00906231626868248, + 0.00723381619900465, + 0.021549316123127937, + 0.018552759662270546, + 0.011740149930119514, + 0.00790036004036665, + 0.010816418565809727, + 0.014854425564408302, + 0.016248147934675217, + 0.016939658671617508, + 0.014490995556116104, + 0.013923360034823418, + 0.024186471477150917, + 0.010357496328651905, + 0.007935957983136177, + 0.008633476682007313, + 0.008207645267248154, + 0.006368768867105246, + 0.006948832888156176, + 0.010418836958706379, + 0.007422707509249449, + 0.013528610579669476, + 0.0064712464809417725, + 0.009528719820082188, + 0.008773247711360455, + 0.009113924577832222, + 0.007003283128142357, + 0.006333976052701473, + 0.007016983814537525, + 0.011456677690148354, + 0.02002105489373207, + 0.022490324452519417, + 0.019925348460674286, + 0.025045381858944893, + 0.012508770450949669, + 0.01327582262456417, + 0.014599169604480267, + 0.018967045471072197, + 0.006641284096986055, + 0.0062147146090865135, + 0.011601012200117111, + 0.010511224158108234, + 0.00881795585155487, + 0.004478071350604296, + 0.008862366899847984, + 0.007087538484483957, + 0.003369744401425123, + 0.008472846820950508, + 0.005003190599381924, + 0.00969403050839901, + 0.005540295969694853, + 0.005007987376302481, + 0.009329685941338539, + 0.0064694578759372234, + 0.0051947226747870445, + 0.0067108068615198135, + 0.005411112681031227, + 0.008161457255482674, + 0.007736586965620518, + 0.008855970576405525, + 0.01091593224555254, + 0.010150856338441372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.04863683879375458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, low-angle, handheld, Steadicam tracking shot closely follows a young woman as she dashes with agility and power through the ruins of an ancient castle, densely covered in vines. Her dark brown hair, worn in a high ponytail, a few sweat-soaked strands plastered to her face in a focused, anxious state. Her lips are pursed, and her eyes scan the obstacle-filled path ahead with a sharp gaze. She wears a fitted charcoal gray overalls, sleeves casually rolled to the elbows, and a pair of worn brown leather hiking boots. Each step is precise and swift over the scattered rubble and broken stone pillars. Golden afternoon sunlight filters through the crumbling Gothic arches and gaping stone walls overhead, slicing through the air thick with ancient dust, creating distinct, Tyndall-effect beams that oscillate with the camera's movement, casting shifting, dappled shadows across the ancient stone pavement. The entire image is imbued with a cinematic quality, with sharp focus and a subtle sense of film grain, perfectly capturing the dynamic beauty and suspenseful tension of this thrilling chase.", + "seed": 13357, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_9/index_3540_seed_13357_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015730244386941195, + "vlm_logprob": -0.015730244386941195, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.015730244386941195, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012603631243109703 + ], + "opposite_logprob_tokens": [ + -4.512603759765625 + ], + "logprob_reward": -0.012603631243109703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012603631243109703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06133301556110382, + 0.027947982773184776, + 0.07382848113775253, + 0.1872052550315857, + 0.03810448199510574, + 0.08375543355941772, + 0.043747078627347946, + 0.1111653596162796, + 0.05835341662168503, + 0.31455957889556885 + ], + "credit_map": [ + 0.007545768283307552, + 0.00786284077912569, + 0.005481356754899025, + 0.007040485739707947, + 0.010694020427763462, + 0.005237373057752848, + 0.008204215206205845, + 0.009266952984035015, + 0.0031836789567023516, + 0.004053728189319372, + 0.003940650727599859, + 0.002724414225667715, + 0.0028824491892009974, + 0.004252979066222906, + 0.002962344791740179, + 0.003947737626731396, + 0.007074164692312479, + 0.010853707790374756, + 0.008416394703090191, + 0.015785474330186844, + 0.0056008524261415005, + 0.007617135997861624, + 0.010419623926281929, + 0.00806113425642252, + 0.0077629717998206615, + 0.005384106654673815, + 0.00799323245882988, + 0.008534750901162624, + 0.007807289250195026, + 0.07204839587211609, + 0.010104137472808361, + 0.06757035851478577, + 0.0031865097116678953, + 0.0028381224256008863, + 0.006020945962518454, + 0.004512961022555828, + 0.006737592164427042, + 0.004198712762445211, + 0.00640779547393322, + 0.004201842937618494, + 0.004174160771071911, + 0.004380400292575359, + 0.0053375838324427605, + 0.007300018798559904, + 0.008715757168829441, + 0.0069688051007688046, + 0.023517366498708725, + 0.023361336439847946, + 0.0030021134298294783, + 0.0049510737881064415, + 0.004831709899008274, + 0.010549500584602356, + 0.004351177252829075, + 0.005709643475711346, + 0.004399007651954889, + 0.005952851381152868, + 0.01255436334758997, + 0.011596721597015858, + 0.01627383567392826, + 0.014078397303819656, + 0.02204986847937107, + 0.011130951344966888, + 0.01308484934270382, + 0.010396365076303482, + 0.007341891527175903, + 0.00970884133130312, + 0.008626147173345089, + 0.009721692651510239, + 0.006489894352853298, + 0.0056577641516923904, + 0.0054848287254571915, + 0.0053223599679768085, + 0.014476640149950981, + 0.04113518074154854, + 0.03443543612957001, + 0.03351100906729698, + 0.042598482221364975, + 0.07645701617002487, + 0.030746497213840485, + 0.041199278086423874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.012603631243109703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072239726781845 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072239726781845, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072239726781845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10236414521932602, + 0.035262580960989, + 0.08089777082204819, + 0.0678028091788292, + 0.10287144035100937, + 0.04247002676129341, + 0.3832584023475647, + 0.07262372970581055, + 0.06475341320037842, + 0.047695644199848175 + ], + "credit_map": [ + 0.01104760356247425, + 0.011339751072227955, + 0.009964331984519958, + 0.012014833278954029, + 0.020587395876646042, + 0.011606646701693535, + 0.013443255797028542, + 0.012360330671072006, + 0.0035935863852500916, + 0.0047616735100746155, + 0.006027303170412779, + 0.0046071335673332214, + 0.0035034127067774534, + 0.004707454703748226, + 0.003303118981420994, + 0.004758899100124836, + 0.0058775548823177814, + 0.010609680786728859, + 0.007567835971713066, + 0.014884570613503456, + 0.007184801157563925, + 0.014382225461304188, + 0.011289086192846298, + 0.009102021344006062, + 0.004475785419344902, + 0.0028267931193113327, + 0.005116963759064674, + 0.005624589044600725, + 0.003714936086907983, + 0.020006582140922546, + 0.007751591969281435, + 0.018285563215613365, + 0.007007582578808069, + 0.007082877680659294, + 0.014283796772360802, + 0.008988350629806519, + 0.013224872760474682, + 0.025517966598272324, + 0.014820144511759281, + 0.011945844627916813, + 0.0036005228757858276, + 0.005795691628009081, + 0.005342245567589998, + 0.004822177346795797, + 0.0048250798135995865, + 0.0056589809246361256, + 0.006106413435190916, + 0.006318916566669941, + 0.03489352762699127, + 0.0365576446056366, + 0.03621388599276543, + 0.03638129308819771, + 0.04747302457690239, + 0.08907457441091537, + 0.05281227082014084, + 0.04985221102833748, + 0.0053239986300468445, + 0.009440510533750057, + 0.010786685161292553, + 0.011920996941626072, + 0.010667387396097183, + 0.009034897200763226, + 0.008927233517169952, + 0.0065220193937420845, + 0.007802915293723345, + 0.006375400349497795, + 0.010972232557833195, + 0.013590624555945396, + 0.009473739191889763, + 0.00452324328944087, + 0.006863411050289869, + 0.005151846446096897, + 0.004426219966262579, + 0.004450217820703983, + 0.007829403504729271, + 0.007426901254802942, + 0.005399139132350683, + 0.007298864424228668, + 0.0032358006574213505, + 0.00762909883633256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.016072239726781845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944311618804932 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944311618804932, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944311618804932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.093992680311203, + 0.08512333780527115, + 0.05505465716123581, + 0.061279766261577606, + 0.04808893799781799, + 0.21863611042499542, + 0.17633198201656342, + 0.14080531895160675, + 0.05699634552001953, + 0.06369081884622574 + ], + "credit_map": [ + 0.009126394055783749, + 0.0077397506684064865, + 0.00849191751331091, + 0.00876076240092516, + 0.023346630856394768, + 0.011864875443279743, + 0.012465094216167927, + 0.012197259813547134, + 0.010029018856585026, + 0.012803705409169197, + 0.015022048726677895, + 0.008761622942984104, + 0.007279265206307173, + 0.008197314105927944, + 0.011535011231899261, + 0.011495358310639858, + 0.00493459589779377, + 0.0061502475291490555, + 0.004353800322860479, + 0.0054230522364377975, + 0.0054856473580002785, + 0.016597017645835876, + 0.004960159305483103, + 0.007150140590965748, + 0.00729501573368907, + 0.0038562624249607325, + 0.007301037199795246, + 0.007386266253888607, + 0.007873517461121082, + 0.008476419374346733, + 0.009761971421539783, + 0.009329278953373432, + 0.0037073506973683834, + 0.004942872561514378, + 0.005987119860947132, + 0.005368705373257399, + 0.006591005716472864, + 0.0073296790942549706, + 0.007711414247751236, + 0.006450791843235493, + 0.01033900585025549, + 0.014173678122460842, + 0.013670753687620163, + 0.015765778720378876, + 0.021823931485414505, + 0.028341326862573624, + 0.0616171695291996, + 0.052904482930898666, + 0.013250269927084446, + 0.018750138580799103, + 0.017032284289598465, + 0.01837843284010887, + 0.02208167500793934, + 0.041154030710458755, + 0.023924704641103745, + 0.021760454401373863, + 0.013403316959738731, + 0.02378382347524166, + 0.016956450417637825, + 0.013611549511551857, + 0.012727228924632072, + 0.025741754099726677, + 0.011411168612539768, + 0.02317003346979618, + 0.0067092375829815865, + 0.005657310597598553, + 0.008986181579530239, + 0.008637496270239353, + 0.008926983922719955, + 0.004922847263514996, + 0.008304184302687645, + 0.004852106794714928, + 0.005860531702637672, + 0.0052457465790212154, + 0.011032535694539547, + 0.008173299953341484, + 0.006049790885299444, + 0.011349772103130817, + 0.004936290439218283, + 0.011042860336601734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.005944311618804932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947392899543047 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947392899543047, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947392899543047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13223163783550262, + 0.14796662330627441, + 0.05403328314423561, + 0.2953253984451294, + 0.03781282901763916, + 0.09562606364488602, + 0.08169493079185486, + 0.0697525143623352, + 0.026292456313967705, + 0.05926430597901344 + ], + "credit_map": [ + 0.01195068284869194, + 0.010933843441307545, + 0.010202819481492043, + 0.012946858070790768, + 0.030424019321799278, + 0.01747797429561615, + 0.019214414060115814, + 0.019081009551882744, + 0.013968953862786293, + 0.01809905655682087, + 0.018577460199594498, + 0.013305479660630226, + 0.01913839392364025, + 0.027606511488556862, + 0.017491107806563377, + 0.019779648631811142, + 0.005804792512208223, + 0.007218512706458569, + 0.00609770230948925, + 0.00797685794532299, + 0.006026184186339378, + 0.007762980181723833, + 0.007708848919719458, + 0.005437402054667473, + 0.011607073247432709, + 0.00761025445535779, + 0.010129069909453392, + 0.006763298995792866, + 0.007146388757973909, + 0.12300661951303482, + 0.014396868646144867, + 0.11466579884290695, + 0.003281253855675459, + 0.004368980880826712, + 0.00412295525893569, + 0.00392268318682909, + 0.005846322048455477, + 0.004943167790770531, + 0.006942070089280605, + 0.004385393112897873, + 0.007268677465617657, + 0.0072257025167346, + 0.008505853824317455, + 0.00979743991047144, + 0.017974093556404114, + 0.016292570158839226, + 0.011349974200129509, + 0.01721174269914627, + 0.0070348698645830154, + 0.008194386027753353, + 0.007332920096814632, + 0.007266818080097437, + 0.009469959884881973, + 0.019392451271414757, + 0.012161929160356522, + 0.010841584764420986, + 0.007430138997733593, + 0.011138255707919598, + 0.008439282886683941, + 0.00971862394362688, + 0.007117564789950848, + 0.011134044267237186, + 0.007008417975157499, + 0.007766188587993383, + 0.002592517761513591, + 0.002729871543124318, + 0.0033988466020673513, + 0.0027757794596254826, + 0.005300845950841904, + 0.0035903984680771828, + 0.0037360915448516607, + 0.0021680998615920544, + 0.004739000927656889, + 0.006764612626284361, + 0.009623710066080093, + 0.011685137636959553, + 0.00678740069270134, + 0.007420724257826805, + 0.005476644262671471, + 0.0067670694552361965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.005947392899543047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing charcoal gray overalls with sleeves rolled to the elbows and brown leather hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of an ancient castle covered in vines with scattered rubble and broken stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature golden afternoon sunlight filtering through arches to create distinct beams of light and dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle handheld Steadicam tracking shot following the woman as she dashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038083646446466446 + ], + "opposite_logprob_tokens": [ + -3.288083553314209 + ], + "logprob_reward": -0.038083646446466446, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038083646446466446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06918208301067352, + 0.04250528663396835, + 0.11859963089227676, + 0.23357772827148438, + 0.07907497882843018, + 0.07478785514831543, + 0.14478562772274017, + 0.05820057913661003, + 0.13805070519447327, + 0.04123552143573761 + ], + "credit_map": [ + 0.006202345713973045, + 0.00908310990780592, + 0.006285039242357016, + 0.007220200728625059, + 0.011030327528715134, + 0.006233088672161102, + 0.008951136842370033, + 0.014176829718053341, + 0.003950717393308878, + 0.008145920000970364, + 0.005684615578502417, + 0.00494666351005435, + 0.0036484526935964823, + 0.0049064671620726585, + 0.0055982815101742744, + 0.0056241643615067005, + 0.010297982953488827, + 0.01722949743270874, + 0.017123593017458916, + 0.024653377011418343, + 0.007661222945898771, + 0.013519234955310822, + 0.01765567809343338, + 0.01045904029160738, + 0.015067570842802525, + 0.009030901826918125, + 0.012212870642542839, + 0.012051012367010117, + 0.011243349872529507, + 0.08230102807283401, + 0.014156794175505638, + 0.07751418650150299, + 0.004895646125078201, + 0.005636773072183132, + 0.009443532675504684, + 0.00868203118443489, + 0.014278305694460869, + 0.008748499676585197, + 0.015492984093725681, + 0.011897191405296326, + 0.005945027340203524, + 0.007181044667959213, + 0.007062842138111591, + 0.006453607697039843, + 0.013745764270424843, + 0.015159265138208866, + 0.00859122071415186, + 0.010649077594280243, + 0.010703202337026596, + 0.017512189224362373, + 0.015238307416439056, + 0.013297258876264095, + 0.017112180590629578, + 0.03159547224640846, + 0.02155432291328907, + 0.017772680148482323, + 0.006014864891767502, + 0.008597378619015217, + 0.008054771460592747, + 0.008375304751098156, + 0.0083466162905097, + 0.005522661376744509, + 0.008625267073512077, + 0.004663706291466951, + 0.01245682593435049, + 0.016674552112817764, + 0.01845950447022915, + 0.016894657164812088, + 0.02381926216185093, + 0.016022691503167152, + 0.022127913311123848, + 0.011595289222896099, + 0.0037195158656686544, + 0.006433623377233744, + 0.004797352012246847, + 0.0072981431148946285, + 0.0029477833304554224, + 0.005473824217915535, + 0.0035124041605740786, + 0.007052870932966471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.038083646446466446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file